feat(FN-3195): document llama.cpp provider setup and onboarding

Docs(FN-3195): adds llama.cpp provider setup and onboarding documentation to the main and dashboard READMEs.

Fusion-Task-Id: FN-3195

.changeset/add-llama-cpp-provider.md

@@ -0,0 +1,5 @@
+---
+"@runfusion/fusion": minor
+---
+
+Add first-class llama.cpp provider support with bundled extension wiring, dashboard status/auth routes, model filtering, and onboarding/settings UI for enabling llama-server models without manual `pi install` steps.

packages/cli/package.json

@@ -87,6 +87,7 @@
     "@fusion/dashboard": "workspace:*",
     "@fusion/engine": "workspace:*",
     "@fusion/pi-claude-cli": "workspace:*",
+    "@fusion/pi-llama-cpp": "workspace:*",
     "@types/node": "^22.0.0",
     "@types/react": "^19.0.0",
     "@vitest/coverage-v8": "^3.1.0",

packages/cli/src/commands/__tests__/llama-cpp-extension.test.ts

@@ -0,0 +1,43 @@
+import { describe, expect, it } from "vitest";
+import {
+  resolveLlamaCppExtension,
+  resolveLlamaCppExtensionPaths,
+} from "../llama-cpp-extension.js";
+
+describe("resolveLlamaCppExtension", () => {
+  it("finds the bundled @fusion/pi-llama-cpp package", () => {
+    const result = resolveLlamaCppExtension();
+    expect(result.status).toBe("ok");
+    if (result.status === "ok") {
+      expect(result.path).toMatch(/pi-llama-cpp[\\/]index\.ts$/);
+      expect(result.packageVersion).toMatch(/^\d+\.\d+\.\d+$/);
+    }
+  });
+});
+
+describe("resolveLlamaCppExtensionPaths", () => {
+  it("returns empty when useLlamaCpp is off", () => {
+    const result = resolveLlamaCppExtensionPaths({});
+    expect(result.paths).toEqual([]);
+    expect(result.warning).toBeUndefined();
+    expect(result.resolution).toBeNull();
+  });
+
+  it("returns extension path when useLlamaCpp is on", () => {
+    const result = resolveLlamaCppExtensionPaths({ useLlamaCpp: true });
+    expect(result.paths).toHaveLength(1);
+    expect(result.paths[0]).toMatch(/pi-llama-cpp[\\/]index\.ts$/);
+    expect(result.resolution?.status).toBe("ok");
+  });
+});
+
+describe("cached resolution roundtrip", () => {
+  it("set/get preserves snapshot", async () => {
+    const { setCachedLlamaCppResolution, getCachedLlamaCppResolution } =
+      await import("../llama-cpp-extension.js");
+    setCachedLlamaCppResolution({ status: "not-installed" });
+    expect(getCachedLlamaCppResolution()).toEqual({ status: "not-installed" });
+    setCachedLlamaCppResolution(null);
+    expect(getCachedLlamaCppResolution()).toBeNull();
+  });
+});

packages/cli/src/commands/daemon.ts

@@ -54,6 +54,11 @@ import {
   resolveDroidCliExtensionPaths,
   setCachedDroidCliResolution,
 } from "./droid-cli-extension.js";
+import {
+  getCachedLlamaCppResolution,
+  resolveLlamaCppExtensionPaths,
+  setCachedLlamaCppResolution,
+} from "./llama-cpp-extension.js";
 import { resolveSelfExtension } from "./self-extension.js";
 import { createReadOnlyAuthFileStorage, mergeAuthStorageReads, wrapAuthStorageWithApiKeyProviders } from "./provider-auth.js";
 import { getCodexCliAuthPath, getFusionAuthPath, getLegacyAuthPaths, getModelRegistryModelsPath, getPackageManagerAgentDir } from "./auth-paths.js";
@@ -473,6 +478,24 @@ export async function runDaemon(opts: DaemonOptions = {}) {
       }
     })();
 
+    const llamaCppPaths = await (async () => {
+      try {
+        const globalSettings = await store.getGlobalSettingsStore().getSettings();
+        const result = resolveLlamaCppExtensionPaths(globalSettings);
+        setCachedLlamaCppResolution(result.resolution);
+        if (result.warning) {
+          console.warn(`[extensions] llama-cpp: ${result.warning}`);
+        }
+        return result.paths;
+      } catch (err) {
+        console.warn(
+          `[extensions] Unable to evaluate useLlamaCpp setting: ${err instanceof Error ? err.message : String(err)}`,
+        );
+        setCachedLlamaCppResolution(null);
+        return [];
+      }
+    })();
+
     // Always prefer Fusion's vendored `@fusion/pi-claude-cli` over any
     // external `pi-claude-cli` install. Drops shadowing externals (e.g. a
     // global `npm install -g pi-claude-cli`) so the upstream's once-and-lock
@@ -492,7 +515,7 @@ export async function runDaemon(opts: DaemonOptions = {}) {
     );
 
     const extensionsResult = await discoverAndLoadExtensions(
-      [...reconciledExtensionPaths, ...droidCliPaths],
+      [...reconciledExtensionPaths, ...droidCliPaths, ...llamaCppPaths],
       cwd,
       join(cwd, ".fusion", "disabled-auto-extension-discovery"),
     );
@@ -586,6 +609,17 @@ export async function runDaemon(opts: DaemonOptions = {}) {
       }
       return { status: r.status, reason: r.reason };
     },
+    getLlamaCppExtensionStatus: () => {
+      const r = getCachedLlamaCppResolution();
+      if (!r) return null;
+      if (r.status === "ok") {
+        return { status: "ok", path: r.path, packageVersion: r.packageVersion };
+      }
+      if (r.status === "not-installed") {
+        return { status: "not-installed" };
+      }
+      return { status: r.status, reason: r.reason };
+    },
     onUseClaudeCliToggled: (_prev, next) => {
       if (!next) return;
       void (async () => {

packages/cli/src/commands/dashboard.ts

@@ -53,6 +53,11 @@ import {
   resolveDroidCliExtensionPaths,
   setCachedDroidCliResolution,
 } from "./droid-cli-extension.js";
+import {
+  getCachedLlamaCppResolution,
+  resolveLlamaCppExtensionPaths,
+  setCachedLlamaCppResolution,
+} from "./llama-cpp-extension.js";
 import { getCachedUpdateStatus, isUpdateCheckEnabled } from "../update-cache.js";
 import { resolveSelfExtension } from "./self-extension.js";
 import { ensureBundledDependencyGraphPluginInstalled } from "../plugins/bundled-plugin-install.js";
@@ -1272,6 +1277,24 @@ export async function runDashboard(port: number, opts: { paused?: boolean; dev?:
       }
     })();
 
+    const llamaCppPaths = await (async () => {
+      try {
+        const globalSettings = await store.getGlobalSettingsStore().getSettings();
+        const result = resolveLlamaCppExtensionPaths(globalSettings);
+        setCachedLlamaCppResolution(result.resolution);
+        if (result.warning) {
+          console.warn(`[extensions] llama-cpp: ${result.warning}`);
+        }
+        return result.paths;
+      } catch (err) {
+        console.warn(
+          `[extensions] Unable to evaluate useLlamaCpp setting: ${err instanceof Error ? err.message : String(err)}`,
+        );
+        setCachedLlamaCppResolution(null);
+        return [];
+      }
+    })();
+
     // Always inject the cli's own extension (`@runfusion/fusion`) so its
     // `fn_*` tools register globally even when the user hasn't run
     // `pi install npm:@runfusion/fusion`. Without this, agent chat with
@@ -1294,6 +1317,7 @@ export async function runDashboard(port: number, opts: { paused?: boolean; dev?:
         ...packageExtensionPaths,
         ...claudeCliPaths,
         ...droidCliPaths,
+        ...llamaCppPaths,
       ],
       cwd,
       join(cwd, ".fusion", "disabled-auto-extension-discovery"),
@@ -1580,6 +1604,17 @@ export async function runDashboard(port: number, opts: { paused?: boolean; dev?:
         }
         return { status: r.status, reason: r.reason };
       },
+      getLlamaCppExtensionStatus: () => {
+        const r = getCachedLlamaCppResolution();
+        if (!r) return null;
+        if (r.status === "ok") {
+          return { status: "ok", path: r.path, packageVersion: r.packageVersion };
+        }
+        if (r.status === "not-installed") {
+          return { status: "not-installed" };
+        }
+        return { status: r.status, reason: r.reason };
+      },
       onUseClaudeCliToggled: (_prev, next) => {
         if (!next) return;
         void (async () => {
@@ -1829,6 +1864,17 @@ export async function runDashboard(port: number, opts: { paused?: boolean; dev?:
         }
         return { status: r.status, reason: r.reason };
       },
+      getLlamaCppExtensionStatus: () => {
+        const r = getCachedLlamaCppResolution();
+        if (!r) return null;
+        if (r.status === "ok") {
+          return { status: "ok", path: r.path, packageVersion: r.packageVersion };
+        }
+        if (r.status === "not-installed") {
+          return { status: "not-installed" };
+        }
+        return { status: r.status, reason: r.reason };
+      },
       onUseClaudeCliToggled: (_prev, next) => {
         if (!next) return;
         void (async () => {

packages/cli/src/commands/llama-cpp-extension.ts

@@ -0,0 +1,114 @@
+import { existsSync, readFileSync } from "node:fs";
+import { createRequire } from "node:module";
+import { dirname, resolve } from "node:path";
+import { fileURLToPath } from "node:url";
+
+const require_ = createRequire(import.meta.url);
+
+export type LlamaCppExtensionResolution =
+  | { status: "ok"; path: string; packageVersion: string }
+  | { status: "not-installed" }
+  | { status: "missing-entry"; reason: string }
+  | { status: "error"; reason: string };
+
+export function resolveLlamaCppExtensionFromModuleUrl(
+  moduleUrl: string,
+): LlamaCppExtensionResolution {
+  let pkgJsonPath: string | undefined;
+
+  const here = dirname(fileURLToPath(moduleUrl));
+  for (const rel of ["pi-llama-cpp", "../pi-llama-cpp", "../../pi-llama-cpp"]) {
+    const candidate = resolve(here, rel, "package.json");
+    if (existsSync(candidate)) {
+      pkgJsonPath = candidate;
+      break;
+    }
+  }
+
+  if (!pkgJsonPath) {
+    try {
+      pkgJsonPath = require_.resolve("@fusion/pi-llama-cpp/package.json");
+    } catch {
+      return { status: "not-installed" };
+    }
+  }
+
+  let pkgJson: { pi?: { extensions?: unknown }; version?: string };
+  try {
+    pkgJson = JSON.parse(readFileSync(pkgJsonPath, "utf-8")) as typeof pkgJson;
+  } catch (err) {
+    return {
+      status: "error",
+      reason: `Failed to read @fusion/pi-llama-cpp package.json: ${err instanceof Error ? err.message : String(err)}`,
+    };
+  }
+
+  const extensions = pkgJson.pi?.extensions;
+  if (!Array.isArray(extensions) || extensions.length === 0) {
+    return {
+      status: "missing-entry",
+      reason: "@fusion/pi-llama-cpp package.json has no pi.extensions array",
+    };
+  }
+
+  const rawEntry = extensions[0];
+  if (typeof rawEntry !== "string" || rawEntry.length === 0) {
+    return {
+      status: "missing-entry",
+      reason: "@fusion/pi-llama-cpp pi.extensions[0] is not a valid path string",
+    };
+  }
+
+  const entryPath = resolve(dirname(pkgJsonPath), rawEntry);
+  if (!existsSync(entryPath)) {
+    return {
+      status: "missing-entry",
+      reason: `@fusion/pi-llama-cpp extension file not found at ${entryPath}`,
+    };
+  }
+
+  return { status: "ok", path: entryPath, packageVersion: pkgJson.version ?? "unknown" };
+}
+
+export function resolveLlamaCppExtension(): LlamaCppExtensionResolution {
+  return resolveLlamaCppExtensionFromModuleUrl(import.meta.url);
+}
+
+export function resolveLlamaCppExtensionPaths(globalSettings: {
+  useLlamaCpp?: unknown;
+}): { paths: string[]; warning?: string; resolution: LlamaCppExtensionResolution | null } {
+  const enabled = globalSettings?.useLlamaCpp === true;
+  if (!enabled) return { paths: [], resolution: null };
+
+  const resolution = resolveLlamaCppExtension();
+  switch (resolution.status) {
+    case "ok":
+      return { paths: [resolution.path], resolution };
+    case "not-installed":
+      return {
+        paths: [],
+        resolution,
+        warning:
+          "useLlamaCpp is on but @fusion/pi-llama-cpp is not installed in node_modules. Run `pnpm install`.",
+      };
+    case "missing-entry":
+    case "error":
+      return { paths: [], resolution, warning: resolution.reason };
+  }
+}
+
+let cachedResolution: LlamaCppExtensionResolution | null = null;
+
+export function setCachedLlamaCppResolution(
+  resolution: LlamaCppExtensionResolution | null,
+): void {
+  cachedResolution = resolution;
+}
+
+export function getCachedLlamaCppResolution(): LlamaCppExtensionResolution | null {
+  return cachedResolution;
+}
+
+export const _testInternals = {
+  moduleUrl: (): string => fileURLToPath(import.meta.url),
+};

packages/cli/src/commands/serve.ts

@@ -57,6 +57,11 @@ import {
   resolveDroidCliExtensionPaths,
   setCachedDroidCliResolution,
 } from "./droid-cli-extension.js";
+import {
+  getCachedLlamaCppResolution,
+  resolveLlamaCppExtensionPaths,
+  setCachedLlamaCppResolution,
+} from "./llama-cpp-extension.js";
 import { resolveSelfExtension } from "./self-extension.js";
 import { registerCustomProviders, reregisterCustomProviders } from "./custom-provider-registry.js";
 import { ensureBundledDependencyGraphPluginInstalled } from "../plugins/bundled-plugin-install.js";
@@ -539,6 +544,24 @@ export async function runServe(
       }
     })();
 
+    const llamaCppPaths = await (async () => {
+      try {
+        const globalSettings = await store.getGlobalSettingsStore().getSettings();
+        const result = resolveLlamaCppExtensionPaths(globalSettings);
+        setCachedLlamaCppResolution(result.resolution);
+        if (result.warning) {
+          console.warn(`[extensions] llama-cpp: ${result.warning}`);
+        }
+        return result.paths;
+      } catch (err) {
+        console.warn(
+          `[extensions] Unable to evaluate useLlamaCpp setting: ${err instanceof Error ? err.message : String(err)}`,
+        );
+        setCachedLlamaCppResolution(null);
+        return [];
+      }
+    })();
+
     // Inject the cli's own extension so fn_* tools register globally without
     // requiring `pi install npm:@runfusion/fusion`.
     const selfExtension = resolveSelfExtension();
@@ -555,6 +578,7 @@ export async function runServe(
         ...packageExtensionPaths,
         ...claudeCliPaths,
         ...droidCliPaths,
+        ...llamaCppPaths,
       ],
       cwd,
       join(cwd, ".fusion", "disabled-auto-extension-discovery"),
@@ -781,6 +805,17 @@ export async function runServe(
       }
       return { status: r.status, reason: r.reason };
     },
+    getLlamaCppExtensionStatus: () => {
+      const r = getCachedLlamaCppResolution();
+      if (!r) return null;
+      if (r.status === "ok") {
+        return { status: "ok", path: r.path, packageVersion: r.packageVersion };
+      }
+      if (r.status === "not-installed") {
+        return { status: "not-installed" };
+      }
+      return { status: r.status, reason: r.reason };
+    },
     onUseClaudeCliToggled: (_prev, next) => {
       if (!next) return; // Toggle-off leaves existing skill symlinks alone.
       void (async () => {

packages/cli/tsup.config.ts

@@ -10,6 +10,8 @@ const piClaudeCliSrc = join(__dirname, "..", "pi-claude-cli");
 const piClaudeCliDest = join(__dirname, "dist", "pi-claude-cli");
 const droidCliSrc = join(__dirname, "..", "droid-cli");
 const droidCliDest = join(__dirname, "dist", "droid-cli");
+const llamaCppSrc = join(__dirname, "..", "pi-llama-cpp");
+const llamaCppDest = join(__dirname, "dist", "pi-llama-cpp");
 const dependencyGraphPluginSrc = join(__dirname, "..", "..", "plugins", "fusion-plugin-dependency-graph");
 const dependencyGraphPluginDest = join(__dirname, "dist", "plugins", "fusion-plugin-dependency-graph");
 const dashboardClientStub = `<!doctype html>
@@ -96,6 +98,21 @@ export default defineConfig({
       );
     }
 
+    if (existsSync(llamaCppDest)) {
+      rmSync(llamaCppDest, { recursive: true, force: true });
+    }
+    if (existsSync(llamaCppSrc)) {
+      mkdirSync(llamaCppDest, { recursive: true });
+      cpSync(join(llamaCppSrc, "index.ts"), join(llamaCppDest, "index.ts"));
+      cpSync(join(llamaCppSrc, "src"), join(llamaCppDest, "src"), { recursive: true });
+      cpSync(join(llamaCppSrc, "package.json"), join(llamaCppDest, "package.json"));
+      console.log("Copied pi-llama-cpp extension to dist/pi-llama-cpp/");
+    } else {
+      console.warn(
+        `WARNING: pi-llama-cpp source not found at ${llamaCppSrc}; useLlamaCpp will not work in the published package.`,
+      );
+    }
+
     if (existsSync(dependencyGraphPluginDest)) {
       rmSync(dependencyGraphPluginDest, { recursive: true, force: true });
     }

packages/core/src/__tests__/use-llama-cpp-settings.test.ts

@@ -0,0 +1,28 @@
+import { describe, expect, it } from "vitest";
+import type { GlobalSettings } from "../types.js";
+import {
+  DEFAULT_GLOBAL_SETTINGS,
+  GLOBAL_SETTINGS_KEYS,
+  isGlobalSettingsKey,
+} from "../settings-schema.js";
+
+describe("useLlamaCpp global setting", () => {
+  it("is included in GLOBAL_SETTINGS_KEYS", () => {
+    expect(GLOBAL_SETTINGS_KEYS).toContain("useLlamaCpp");
+  });
+
+  it("defaults to undefined", () => {
+    expect(DEFAULT_GLOBAL_SETTINGS.useLlamaCpp).toBeUndefined();
+  });
+
+  it("is recognized by isGlobalSettingsKey", () => {
+    expect(isGlobalSettingsKey("useLlamaCpp")).toBe(true);
+  });
+
+  it("accepts boolean values in GlobalSettings", () => {
+    const enabled: GlobalSettings = { useLlamaCpp: true };
+    const disabled: GlobalSettings = { useLlamaCpp: false };
+    expect(enabled.useLlamaCpp).toBe(true);
+    expect(disabled.useLlamaCpp).toBe(false);
+  });
+});

packages/core/src/settings-schema.ts

@@ -53,6 +53,7 @@ export const DEFAULT_GLOBAL_SETTINGS = {
   modelOnboardingComplete: undefined,
   useClaudeCli: undefined,
   useDroidCli: undefined,
+  useLlamaCpp: undefined,
   // Global baseline lanes for per-role model selection
   executionGlobalProvider: undefined,
   executionGlobalModelId: undefined,

packages/core/src/types.ts

@@ -1403,6 +1403,13 @@ export interface GlobalSettings {
    *  by the dashboard auth toggle. Setting this field explicitly (true/false)
    *  always wins. */
   useDroidCli?: boolean;
+  /** When true, enable llama.cpp model-provider support (provider ID: `llama-server`)
+   *  via Fusion's bundled `@fusion/pi-llama-cpp` extension.
+   *
+   *  When left undefined, llama.cpp routing stays disabled unless explicitly enabled
+   *  by the dashboard auth toggle. Setting this field explicitly (true/false)
+   *  always wins. */
+  useLlamaCpp?: boolean;
   /** Global baseline AI model provider for task execution (executor agent).
    *  This is the global lane that project-level `executionProvider` can override.
    *  Must be set together with `executionGlobalModelId`. Falls back to

packages/dashboard/app/api/legacy.ts

@@ -1317,6 +1317,23 @@ export interface DroidCliStatus {
   ready: boolean;
 }
 
+export interface LlamaCppStatus {
+  enabled: boolean;
+  extension: {
+    status: "ok" | "not-installed" | "missing-entry" | "error";
+    path?: string;
+    packageVersion?: string;
+    reason?: string;
+  } | null;
+  ready: boolean;
+  server: {
+    available: boolean;
+    url: string;
+    hasApiKey: boolean;
+    reason?: string;
+  };
+}
+
 /** Probe the local Claude CLI binary + setting + extension state. */
 export function fetchClaudeCliStatus(): Promise<ClaudeCliStatus> {
   return api<ClaudeCliStatus>("/providers/claude-cli/status");
@@ -1368,6 +1385,11 @@ export function fetchDroidCliStatus(): Promise<DroidCliStatus> {
   return api<DroidCliStatus>("/providers/droid-cli/status");
 }
 
+/** Probe llama.cpp server + setting + extension state. */
+export function fetchLlamaCppStatus(): Promise<LlamaCppStatus> {
+  return api<LlamaCppStatus>("/providers/llama-cpp/status");
+}
+
 // --- Runtime Provider Status Types ---
 
 export interface RuntimeBinaryStatus {
@@ -1627,6 +1649,16 @@ export function setDroidCliEnabled(
   });
 }
 
+/** Enable or disable the llama.cpp provider. */
+export function setLlamaCppEnabled(
+  enabled: boolean,
+): Promise<{ enabled: boolean; restartRequired: boolean }> {
+  return api<{ enabled: boolean; restartRequired: boolean }>("/auth/llama-cpp", {
+    method: "POST",
+    body: JSON.stringify({ enabled }),
+  });
+}
+
 export interface CustomProvider {
   id: string;
   name: string;

packages/dashboard/app/components/LlamaCppProviderCard.css

@@ -0,0 +1,19 @@
+.llama-cpp-provider-card {
+  display: flex;
+  flex-direction: column;
+  gap: var(--space-sm);
+}
+
+.llama-cpp-status {
+  color: var(--text-muted);
+}
+
+.llama-cpp-status--ok {
+  color: var(--color-success);
+}
+
+@media (max-width: 768px) {
+  .llama-cpp-provider-card {
+    gap: var(--space-xs);
+  }
+}

packages/dashboard/app/components/LlamaCppProviderCard.tsx

@@ -0,0 +1,98 @@
+import { useCallback, useEffect, useRef, useState } from "react";
+import { Loader2 } from "lucide-react";
+import { fetchLlamaCppStatus, setLlamaCppEnabled, type LlamaCppStatus } from "../api";
+import { ProviderIcon } from "./ProviderIcon";
+import "./LlamaCppProviderCard.css";
+
+interface LlamaCppProviderCardProps {
+  authenticated: boolean;
+  onToggled?: (nextEnabled: boolean) => void;
+  compact?: boolean;
+}
+
+export function LlamaCppProviderCard({ authenticated, onToggled, compact = false }: LlamaCppProviderCardProps) {
+  const [status, setStatus] = useState<LlamaCppStatus | null>(null);
+  const [busy, setBusy] = useState<"enabling" | "disabling" | "testing" | null>(null);
+  const mountedRef = useRef(true);
+
+  useEffect(() => {
+    mountedRef.current = true;
+    return () => {
+      mountedRef.current = false;
+    };
+  }, []);
+
+  const refresh = useCallback(async () => {
+    const next = await fetchLlamaCppStatus();
+    if (mountedRef.current) setStatus(next);
+    return next;
+  }, []);
+
+  useEffect(() => {
+    void refresh();
+  }, [refresh]);
+
+  const handleToggle = useCallback(async (next: boolean) => {
+    setBusy(next ? "enabling" : "disabling");
+    try {
+      const result = await setLlamaCppEnabled(next);
+      onToggled?.(result.enabled);
+      await refresh();
+    } finally {
+      if (mountedRef.current) setBusy(null);
+    }
+  }, [onToggled, refresh]);
+
+  const handleTest = useCallback(async () => {
+    setBusy("testing");
+    try {
+      await refresh();
+    } finally {
+      if (mountedRef.current) setBusy(null);
+    }
+  }, [refresh]);
+
+  const enabled = status?.enabled ?? authenticated;
+  const serverAvailable = status?.server.available ?? false;
+
+  const content = (
+    <>
+      <div className="auth-provider-info">
+        <ProviderIcon provider="llama-cpp" size={compact ? "sm" : "md"} />
+        <strong>llama.cpp — via HTTP server</strong>
+      </div>
+      <small className={`llama-cpp-status${status?.ready ? " llama-cpp-status--ok" : ""}`}>
+        {!status
+          ? "Probing llama.cpp server…"
+          : status.server.available
+            ? `Server reachable at ${status.server.url}`
+            : `Server unavailable: ${status.server.reason ?? "not reachable"}`}
+      </small>
+      <div className="auth-provider-cli-actions">
+        <button type="button" className="btn btn-sm" onClick={() => void handleTest()} disabled={busy !== null}>
+          {busy === "testing" ? <><Loader2 size={12} className="animate-spin" />Testing…</> : "Test"}
+        </button>
+        {enabled ? (
+          <button type="button" className="btn btn-sm" onClick={() => void handleToggle(false)} disabled={busy !== null}>
+            {busy === "disabling" ? "Disabling…" : "Disable"}
+          </button>
+        ) : (
+          <button
+            type="button"
+            className="btn btn-primary btn-sm"
+            onClick={() => void handleToggle(true)}
+            disabled={busy !== null || !serverAvailable}
+          >
+            {busy === "enabling" ? "Enabling…" : "Enable"}
+          </button>
+        )}
+      </div>
+    </>
+  );
+
+  if (compact) {
+    return <div className="auth-provider-card auth-provider-card--cli llama-cpp-provider-card" data-testid="llama-cpp-provider-card">{content}</div>;
+  }
+
+  return <div className="onboarding-provider-card llama-cpp-provider-card" data-testid="llama-cpp-provider-card">{content}</div>;
+}

packages/dashboard/app/components/ModelOnboardingModal.tsx

@@ -24,6 +24,7 @@ import { CustomModelDropdown } from "./CustomModelDropdown";
 import { ProviderIcon } from "./ProviderIcon";
 import { ClaudeCliProviderCard } from "./ClaudeCliProviderCard";
 import { DroidCliProviderCard } from "./DroidCliProviderCard";
+import { LlamaCppProviderCard } from "./LlamaCppProviderCard";
 import { LoginInstructions } from "./LoginInstructions";
 import { OAuthManualCodeForm } from "./OAuthManualCodeForm";
 import { OnboardingDisclosure } from "./OnboardingDisclosure";
@@ -199,6 +200,7 @@ const ONBOARDING_CURATED_PROVIDER_FAMILY_ORDER = [
   "anthropic",
   "claude-cli",
   "droid-cli",
+  "llama-cpp",
   "openai-codex",
   "gemini",
   "minimax",
@@ -210,6 +212,7 @@ const ONBOARDING_PROVIDER_FAMILY_ALIASES: Record<string, (typeof ONBOARDING_CURA
   anthropic: "anthropic",
   "claude-cli": "claude-cli",
   "droid-cli": "droid-cli",
+  "llama-cpp": "llama-cpp",
   "openai-codex": "openai-codex",
   google: "gemini",
   gemini: "gemini",
@@ -1769,6 +1772,18 @@ export function ModelOnboardingModal({
       );
     }
 
+    if (provider.id === "llama-cpp" && provider.type === "cli") {
+      return (
+        <LlamaCppProviderCard
+          key={provider.id}
+          authenticated={provider.authenticated}
+          onToggled={() => {
+            void loadAuthStatus();
+          }}
+        />
+      );
+    }
+
     if (provider.id === "claude-cli" && provider.type === "cli") {
       return (
         <ClaudeCliProviderCard

packages/dashboard/app/components/ProviderIcon.tsx

@@ -1,5 +1,9 @@
 import { Cpu } from "lucide-react";
 
+function LlamaCppIcon({ size, color, label = "llama.cpp" }: { size: number; color: string; label?: string }) {
+  return <Cpu size={size} color={color} aria-label={label} data-testid="llama-cpp-icon" />;
+}
+
 export interface ProviderIconProps {
   provider: string;
   size?: "sm" | "md" | "lg";
@@ -653,6 +657,8 @@ const providerConfig: Record<
   "claude-cli": { component: ClaudeCliIcon, color: "var(--provider-anthropic)", label: "Anthropic — via Claude CLI" },
   "pi-claude-cli": { component: ClaudeCliIcon, color: "var(--provider-anthropic)", label: "Anthropic — via Claude CLI" },
   "droid-cli": { component: DroidCliIcon, color: "var(--provider-openai)", label: "Factory AI — via Droid CLI" },
+  "llama-cpp": { component: LlamaCppIcon, color: "var(--provider-ollama)", label: "llama.cpp" },
+  "llama-server": { component: LlamaCppIcon, color: "var(--provider-ollama)", label: "llama.cpp" },
 
   openai: { component: OpenAIIcon, color: "var(--provider-openai)" },
   "openai-codex": { component: OpenAIIcon, color: "var(--provider-openai)", label: "OpenAI Codex" }, // OpenAI alias

packages/dashboard/app/components/SettingsModal.tsx

@@ -28,6 +28,7 @@ const PiExtensionsManager = lazy(() => import("./PiExtensionsManager").then((m)
 import { ClaudeCliProviderCard } from "./ClaudeCliProviderCard";
 import { CliBinaryPanel } from "./CliBinaryPanel";
 import { DroidCliProviderCard } from "./DroidCliProviderCard";
+import { LlamaCppProviderCard } from "./LlamaCppProviderCard";
 import { HermesRuntimeCard } from "./HermesRuntimeCard";
 import { OpenClawRuntimeCard } from "./OpenClawRuntimeCard";
 import { PaperclipRuntimeCard } from "./PaperclipRuntimeCard";
@@ -5033,6 +5034,7 @@ export function SettingsModal({
         // auth state (Authenticated when signed in, Available otherwise).
         const claudeCliProvider = cliAuthProviders.find((p) => p.id === "claude-cli");
         const droidCliProvider = cliAuthProviders.find((p) => p.id === "droid-cli");
+        const llamaCppProvider = cliAuthProviders.find((p) => p.id === "llama-cpp");
         const hasDroidPluginSlot = getSlotsForId("settings-provider-card").some(
           (entry) => entry.pluginId === "fusion-plugin-droid-runtime",
         );
@@ -5054,14 +5056,25 @@ export function SettingsModal({
             }}
           />
         ) : null;
+        const llamaCppCard = llamaCppProvider ? (
+          <LlamaCppProviderCard
+            compact
+            authenticated={llamaCppProvider.authenticated}
+            onToggled={() => {
+              void loadAuthStatus();
+            }}
+          />
+        ) : null;
         const showAuthenticatedGroup =
           authenticatedProviders.length > 0
           || (claudeCliProvider?.authenticated ?? false)
-          || ((droidCliProvider?.authenticated ?? false) && !hasDroidPluginSlot);
+          || ((droidCliProvider?.authenticated ?? false) && !hasDroidPluginSlot)
+          || (llamaCppProvider?.authenticated ?? false);
         const showAvailableGroup =
           unauthenticatedProviders.length > 0
           || (claudeCliProvider && !claudeCliProvider.authenticated)
-          || (droidCliProvider && !droidCliProvider.authenticated && !hasDroidPluginSlot);
+          || (droidCliProvider && !droidCliProvider.authenticated && !hasDroidPluginSlot)
+          || (llamaCppProvider && !llamaCppProvider.authenticated);
         return (
           <>
             <h4 className="settings-section-heading">Authentication</h4>
@@ -5085,6 +5098,7 @@ export function SettingsModal({
                   <div className="auth-group-label">Authenticated</div>
                   {claudeCliProvider?.authenticated && claudeCliCard}
                   {droidCliProvider?.authenticated && droidCliCard}
+                  {llamaCppProvider?.authenticated && llamaCppCard}
                   {authenticatedProviders.map((provider) => (
                     <div key={provider.id} className="auth-provider-card auth-provider-card--authenticated">
                       <div className="auth-provider-header">
@@ -5179,6 +5193,7 @@ export function SettingsModal({
                   <div className="auth-group-label">Available</div>
                   {claudeCliProvider && !claudeCliProvider.authenticated && claudeCliCard}
                   {droidCliProvider && !droidCliProvider.authenticated && droidCliCard}
+                  {llamaCppProvider && !llamaCppProvider.authenticated && llamaCppCard}
                   {unauthenticatedProviders.map((provider) => (
                     <div key={provider.id} className="auth-provider-card">
                       <div className="auth-provider-header">

packages/dashboard/app/components/__tests__/LlamaCppProviderCard.test.tsx

@@ -0,0 +1,30 @@
+import { fireEvent, render, screen, waitFor } from "@testing-library/react";
+import { beforeEach, describe, expect, it, vi } from "vitest";
+import { LlamaCppProviderCard } from "../LlamaCppProviderCard";
+
+const fetchLlamaCppStatus = vi.fn();
+const setLlamaCppEnabled = vi.fn();
+
+vi.mock("../../api", () => ({
+  fetchLlamaCppStatus: (...args: unknown[]) => fetchLlamaCppStatus(...args),
+  setLlamaCppEnabled: (...args: unknown[]) => setLlamaCppEnabled(...args),
+}));
+
+describe("LlamaCppProviderCard", () => {
+  beforeEach(() => {
+    fetchLlamaCppStatus.mockResolvedValue({
+      enabled: false,
+      ready: false,
+      extension: { status: "ok" },
+      server: { available: true, url: "http://127.0.0.1:8080", hasApiKey: false },
+    });
+    setLlamaCppEnabled.mockResolvedValue({ enabled: true, restartRequired: false });
+  });
+
+  it("renders and enables provider", async () => {
+    render(<LlamaCppProviderCard authenticated={false} />);
+    await waitFor(() => expect(fetchLlamaCppStatus).toHaveBeenCalled());
+    fireEvent.click(screen.getByRole("button", { name: "Enable" }));
+    await waitFor(() => expect(setLlamaCppEnabled).toHaveBeenCalledWith(true));
+  });
+});

packages/dashboard/app/components/__tests__/ProviderIcon.test.tsx

@@ -95,6 +95,13 @@ describe("ProviderIcon", () => {
     expect(screen.getByLabelText("Ollama")).toBeInTheDocument();
   });
 
+  it("renders llama.cpp icon aliases", () => {
+    const { rerender } = render(<ProviderIcon provider="llama-cpp" />);
+    expect(screen.getByTestId("llama-cpp-icon")).toBeInTheDocument();
+    rerender(<ProviderIcon provider="llama-server" />);
+    expect(screen.getByTestId("llama-cpp-icon")).toBeInTheDocument();
+  });
+
   it("renders Cpu icon as fallback for unknown providers", () => {
     render(<ProviderIcon provider="unknown" />);
     // Cpu icon from lucide-react renders as an svg without our custom data-testid

packages/dashboard/app/styles.css

@@ -258,6 +258,7 @@ html {
   --provider-groq: #f55036;
   --provider-vercel: var(--text);
   --provider-droid-cli: var(--text);
+  --provider-ollama: #d4a27f;
   /* Runtime-plugin marks. */
   --provider-hermes: #d4961c;
   --provider-openclaw: #ff4f40;

packages/dashboard/src/__tests__/register-model-routes-llama-cpp.test.ts

@@ -0,0 +1,72 @@
+import { describe, expect, it, vi } from "vitest";
+import type { Router } from "express";
+import { registerModelRoutes } from "../routes/register-model-routes.js";
+
+function setup(useLlamaCpp?: boolean) {
+  const getHandlers = new Map<string, (req: unknown, res: { json: (body: unknown) => void }) => Promise<void>>();
+  const router = {
+    get: vi.fn((path: string, handler: (req: unknown, res: { json: (body: unknown) => void }) => Promise<void>) => {
+      getHandlers.set(path, handler);
+    }),
+  } as unknown as Router;
+
+  const store = {
+    getGlobalSettingsStore: () => ({
+      getSettings: vi.fn().mockResolvedValue({ useLlamaCpp }),
+    }),
+    getSettingsFast: vi.fn().mockResolvedValue({}),
+  };
+
+  const runtimeLogger = {
+    child: vi.fn(() => ({ warn: vi.fn() })),
+  };
+
+  const modelRegistry = {
+    refresh: vi.fn(),
+    getAvailable: vi.fn(() => [
+      { provider: "llama-server", id: "llama3", name: "Llama 3", reasoning: true, contextWindow: 128000 },
+      { provider: "openai", id: "gpt-5", name: "GPT-5", reasoning: true, contextWindow: 128000 },
+    ]),
+  };
+
+  registerModelRoutes({
+    router,
+    store: store as never,
+    runtimeLogger: runtimeLogger as never,
+    options: { modelRegistry } as never,
+  } as never);
+
+  return { handler: getHandlers.get("/models")! };
+}
+
+describe("registerModelRoutes llama-server filter", () => {
+  it("filters llama-server models when useLlamaCpp is false", async () => {
+    const { handler } = setup(false);
+    const json = vi.fn();
+
+    await handler({}, { json });
+
+    const response = json.mock.calls[0][0] as { models: Array<{ provider: string }> };
+    expect(response.models.some((model) => model.provider === "llama-server")).toBe(false);
+  });
+
+  it("includes llama-server models when useLlamaCpp is true", async () => {
+    const { handler } = setup(true);
+    const json = vi.fn();
+
+    await handler({}, { json });
+
+    const response = json.mock.calls[0][0] as { models: Array<{ provider: string }> };
+    expect(response.models.some((model) => model.provider === "llama-server")).toBe(true);
+  });
+
+  it("filters llama-server models when useLlamaCpp is unset", async () => {
+    const { handler } = setup(undefined);
+    const json = vi.fn();
+
+    await handler({}, { json });
+
+    const response = json.mock.calls[0][0] as { models: Array<{ provider: string }> };
+    expect(response.models.some((model) => model.provider === "llama-server")).toBe(false);
+  });
+});

packages/dashboard/src/__tests__/routes-auth.test.ts

@@ -31,6 +31,7 @@ import { SESSION_CLEANUP_DEFAULT_MAX_AGE_MS } from "../ai-session-store.js";
 import * as usageModule from "../usage.js";
 import * as claudeCliProbeModule from "../claude-cli-probe.js";
 import * as droidCliProbeModule from "../droid-cli-probe.js";
+import * as llamaCppProbeModule from "../llama-cpp-probe.js";
 import * as projectStoreResolver from "../project-store-resolver.js";
 import * as terminalServiceModule from "../terminal-service.js";
 import { get as performGet, request as performRequest } from "../test-request.js";
@@ -562,7 +563,7 @@ describe("GET /auth/status", () => {
     expect(res.status).toBe(200);
     // Filter out synthetic CLI providers — they have dedicated route tests.
     // Structural assertions here are about OAuth + API-key paths only.
-    const providers = res.body.providers.filter((p: any) => p.id !== "claude-cli" && p.id !== "droid-cli");
+    const providers = res.body.providers.filter((p: any) => p.id !== "claude-cli" && p.id !== "droid-cli" && p.id !== "llama-cpp");
     expect(providers).toEqual([
       { id: "anthropic", name: "Anthropic", authenticated: true, type: "oauth", loginInProgress: false },
       { id: "openrouter", name: "OpenRouter", authenticated: false, type: "api_key" },
@@ -587,7 +588,7 @@ describe("GET /auth/status", () => {
     const res = await GET(buildApp(), "/api/auth/status");
 
     expect(res.status).toBe(200);
-    const providers = res.body.providers.filter((p: any) => p.id !== "claude-cli" && p.id !== "droid-cli");
+    const providers = res.body.providers.filter((p: any) => p.id !== "claude-cli" && p.id !== "droid-cli" && p.id !== "llama-cpp");
     expect(providers).toEqual([
       { id: "anthropic", name: "Anthropic", authenticated: true, type: "oauth", loginInProgress: false },
       { id: "github-copilot", name: "GitHub Copilot", authenticated: false, type: "oauth", loginInProgress: false },
@@ -3143,3 +3144,182 @@ describe("Pause/Unpause endpoints", () => {
 });
 
 // --- GitHub Import route tests ---
+
+describe("llama.cpp auth routes", () => {
+  let store: TaskStore;
+  let authStorage: AuthStorageLike;
+
+  function buildApp(options?: Parameters<typeof createApiRoutes>[1]) {
+    const app = express();
+    app.use(express.json());
+    app.use("/api", createApiRoutes(store, { authStorage, ...options }));
+    return app;
+  }
+
+  beforeEach(() => {
+    store = createMockStore({
+      updateGlobalSettings: vi.fn().mockResolvedValue({ useLlamaCpp: true }),
+      getGlobalSettingsStore: vi.fn().mockReturnValue({
+        ...createMockGlobalSettingsStore(),
+        getSettings: vi.fn().mockResolvedValue({ useLlamaCpp: false }),
+      }),
+    });
+    authStorage = createMockAuthStorage();
+    vi.spyOn(llamaCppProbeModule, "probeLlamaCpp").mockResolvedValue({
+      reachable: true,
+      url: "http://127.0.0.1:8080",
+      hasApiKey: false,
+    });
+  });
+
+  it("enables llama.cpp when probe passes", async () => {
+    const res = await REQUEST(buildApp(), "POST", "/api/auth/llama-cpp", JSON.stringify({ enabled: true }), {
+      "content-type": "application/json",
+    });
+    expect(res.status).toBe(200);
+    expect(res.body).toEqual({ enabled: true, restartRequired: false });
+    expect(store.updateGlobalSettings).toHaveBeenCalledWith({ useLlamaCpp: true });
+  });
+
+  it("returns 400 when enabling with unreachable server", async () => {
+    vi.spyOn(llamaCppProbeModule, "probeLlamaCpp").mockResolvedValue({
+      reachable: false,
+      url: "http://127.0.0.1:8080",
+      hasApiKey: false,
+      reason: "llama.cpp server did not return a healthy response",
+    });
+
+    const res = await REQUEST(buildApp(), "POST", "/api/auth/llama-cpp", JSON.stringify({ enabled: true }), {
+      "content-type": "application/json",
+    });
+
+    expect(res.status).toBe(400);
+    expect(res.body.error).toContain("Cannot enable llama.cpp routing");
+  });
+
+  it("disabling works without probing the server", async () => {
+    const probeSpy = vi.spyOn(llamaCppProbeModule, "probeLlamaCpp");
+
+    const res = await REQUEST(buildApp(), "POST", "/api/auth/llama-cpp", JSON.stringify({ enabled: false }), {
+      "content-type": "application/json",
+    });
+
+    expect(res.status).toBe(200);
+    expect(probeSpy).not.toHaveBeenCalled();
+  });
+
+  it("returns 400 for non-boolean enabled", async () => {
+    const res = await REQUEST(buildApp(), "POST", "/api/auth/llama-cpp", JSON.stringify({ enabled: "yes" }), {
+      "content-type": "application/json",
+    });
+
+    expect(res.status).toBe(400);
+  });
+
+  it("returns llama.cpp provider status including API key flag", async () => {
+    vi.spyOn(llamaCppProbeModule, "probeLlamaCpp").mockResolvedValue({
+      reachable: true,
+      url: "http://127.0.0.1:8080",
+      hasApiKey: true,
+    });
+
+    const res = await GET(buildApp(), "/api/providers/llama-cpp/status");
+    expect(res.status).toBe(200);
+    expect(res.body.server.url).toBe("http://127.0.0.1:8080");
+    expect(res.body.server.hasApiKey).toBe(true);
+    expect(res.body.ready).toBe(false);
+  });
+
+  it("marks llama.cpp status not ready when extension resolution fails", async () => {
+    const res = await GET(
+      buildApp({
+        getLlamaCppExtensionStatus: () => ({ status: "error", reason: "extension failed" }),
+      } as Parameters<typeof createApiRoutes>[1]),
+      "/api/providers/llama-cpp/status",
+    );
+
+    expect(res.status).toBe(200);
+    expect(res.body.ready).toBe(false);
+    expect(res.body.extension.status).toBe("error");
+  });
+
+  it("GET /auth/status includes llama-cpp provider with cli type", async () => {
+    store.getGlobalSettingsStore = vi.fn().mockReturnValue({
+      ...createMockGlobalSettingsStore(),
+      getSettings: vi.fn().mockResolvedValue({ useLlamaCpp: true }),
+    });
+
+    const res = await GET(buildApp(), "/api/auth/status");
+    expect(res.status).toBe(200);
+    expect(res.body.providers).toEqual(
+      expect.arrayContaining([
+        expect.objectContaining({
+          id: "llama-cpp",
+          name: "llama.cpp — via HTTP server",
+          type: "cli",
+          authenticated: true,
+        }),
+      ]),
+    );
+  });
+
+  it("GET /auth/status marks llama-cpp unauthenticated when extension status is not ok", async () => {
+    store.getGlobalSettingsStore = vi.fn().mockReturnValue({
+      ...createMockGlobalSettingsStore(),
+      getSettings: vi.fn().mockResolvedValue({ useLlamaCpp: true }),
+    });
+
+    const res = await GET(
+      buildApp({ getLlamaCppExtensionStatus: () => ({ status: "error", reason: "bad ext" }) } as Parameters<
+        typeof createApiRoutes
+      >[1]),
+      "/api/auth/status",
+    );
+
+    expect(res.status).toBe(200);
+    expect(res.body.providers).toEqual(
+      expect.arrayContaining([
+        expect.objectContaining({
+          id: "llama-cpp",
+          authenticated: false,
+          type: "cli",
+        }),
+      ]),
+    );
+  });
+
+  it("fires onUseLlamaCppToggled hook on transition", async () => {
+    const onUseLlamaCppToggled = vi.fn();
+
+    const res = await REQUEST(
+      buildApp({ onUseLlamaCppToggled } as Parameters<typeof createApiRoutes>[1]),
+      "POST",
+      "/api/auth/llama-cpp",
+      JSON.stringify({ enabled: true }),
+      { "Content-Type": "application/json" },
+    );
+
+    expect(res.status).toBe(200);
+    expect(onUseLlamaCppToggled).toHaveBeenCalledWith(false, true);
+  });
+
+  it("PUT /settings/global with useLlamaCpp fires onUseLlamaCppToggled", async () => {
+    const onUseLlamaCppToggled = vi.fn();
+    store.updateGlobalSettings = vi.fn().mockResolvedValue({ useLlamaCpp: true });
+    store.getGlobalSettingsStore = vi.fn().mockReturnValue({
+      ...createMockGlobalSettingsStore(),
+      getSettings: vi.fn().mockResolvedValue({ useLlamaCpp: false, useClaudeCli: false, useDroidCli: false }),
+    });
+
+    const res = await REQUEST(
+      buildApp({ onUseLlamaCppToggled } as Parameters<typeof createApiRoutes>[1]),
+      "PUT",
+      "/api/settings/global",
+      JSON.stringify({ useLlamaCpp: true }),
+      { "Content-Type": "application/json" },
+    );
+
+    expect(res.status).toBe(200);
+    expect(onUseLlamaCppToggled).toHaveBeenCalledWith(false, true);
+  });
+});

packages/dashboard/src/llama-cpp-probe.ts

@@ -0,0 +1,78 @@
+import { readFile } from "node:fs/promises";
+import { join } from "node:path";
+
+const DEFAULT_LLAMA_SERVER_URL = "http://127.0.0.1:8080";
+
+type LlamaProjectConfig = { url?: string };
+type LlamaGlobalConfig = { llamaServerUrl?: string };
+type LlamaAuthConfig = Record<string, { key?: string } | undefined>;
+
+export interface LlamaCppProbeStatus {
+  reachable: boolean;
+  url: string;
+  hasApiKey: boolean;
+  reason?: string;
+}
+
+async function readJson<T>(path: string): Promise<T | null> {
+  try {
+    const raw = await readFile(path, "utf-8");
+    return JSON.parse(raw) as T;
+  } catch {
+    return null;
+  }
+}
+
+function normalizeUrl(url: string): string {
+  return url.trim().replace(/\/+$/, "");
+}
+
+async function resolveLlamaServerUrl(cwd: string): Promise<string> {
+  const projectCfg = await readJson<LlamaProjectConfig>(join(cwd, ".pi", "llama-server.json"));
+  if (projectCfg?.url) return normalizeUrl(projectCfg.url);
+
+  const envUrl = process.env.LLAMA_SERVER_URL;
+  if (envUrl) return normalizeUrl(envUrl);
+
+  const globalCfg = await readJson<LlamaGlobalConfig>(
+    join(process.env.HOME ?? ".", ".pi", "agent", "settings.json"),
+  );
+  if (globalCfg?.llamaServerUrl) return normalizeUrl(globalCfg.llamaServerUrl);
+
+  return DEFAULT_LLAMA_SERVER_URL;
+}
+
+async function resolveLlamaServerApiKey(): Promise<string | undefined> {
+  const authCfg = await readJson<LlamaAuthConfig>(join(process.env.HOME ?? ".", ".pi", "agent", "auth.json"));
+  const key = authCfg?.["llama-server"]?.key?.trim();
+  return key ? key : undefined;
+}
+
+async function isLlamaServerReady(url: string, apiKey?: string): Promise<boolean> {
+  try {
+    const response = await fetch(`${url}/health`, {
+      headers: apiKey ? { Authorization: `Bearer ${apiKey}` } : undefined,
+    });
+    if (!response.ok) {
+      return false;
+    }
+    const payload = (await response.json()) as { status?: string };
+    return payload.status === "ok";
+  } catch {
+    return false;
+  }
+}
+
+export async function probeLlamaCpp(options: { cwd?: string } = {}): Promise<LlamaCppProbeStatus> {
+  const cwd = options.cwd ?? process.cwd();
+  const url = await resolveLlamaServerUrl(cwd);
+  const apiKey = await resolveLlamaServerApiKey();
+  const reachable = await isLlamaServerReady(url, apiKey);
+
+  return {
+    reachable,
+    url,
+    hasApiKey: Boolean(apiKey),
+    reason: reachable ? undefined : "llama.cpp server did not return a healthy response",
+  };
+}

packages/dashboard/src/routes/register-auth-routes.ts

@@ -1,6 +1,7 @@
 import { isGhAvailable, isGhAuthenticated } from "@fusion/core";
 import { probeClaudeCli } from "../claude-cli-probe.js";
 import { probeDroidCli } from "../droid-cli-probe.js";
+import { probeLlamaCpp } from "../llama-cpp-probe.js";
 import { ApiError, badRequest, conflict } from "../api-error.js";
 import { clearUsageCache } from "../usage.js";
 import { invalidateAllGlobalSettingsCaches } from "../project-store-resolver.js";
@@ -283,6 +284,26 @@ export const registerAuthRoutes: ApiRouteRegistrar = (ctx) => {
         });
       }
 
+      // Inject synthetic llama.cpp provider.
+      if (store) {
+        let llamaEnabled = false;
+        try {
+          const globalSettings = await store.getGlobalSettingsStore().getSettings();
+          llamaEnabled = globalSettings.useLlamaCpp === true;
+        } catch {
+          // Best-effort
+        }
+        const llamaExtension = options?.getLlamaCppExtensionStatus?.() ?? null;
+        const extensionOk = llamaExtension === null || llamaExtension.status === "ok";
+        const probe = await probeLlamaCpp();
+        providers.push({
+          id: "llama-cpp",
+          name: "llama.cpp — via HTTP server",
+          authenticated: llamaEnabled && probe.reachable && extensionOk,
+          type: "cli" as const,
+        });
+      }
+
       const ghCli = {
         available: isGhAvailable(),
         authenticated: isGhAuthenticated(),
@@ -517,6 +538,93 @@ export const registerAuthRoutes: ApiRouteRegistrar = (ctx) => {
     }
   });
 
+  router.post("/auth/llama-cpp", async (req, res) => {
+    try {
+      if (!store) {
+        throw new ApiError(500, "Settings store unavailable");
+      }
+      const enabled = req.body?.enabled;
+      if (typeof enabled !== "boolean") {
+        throw badRequest("enabled must be a boolean");
+      }
+
+      if (enabled) {
+        const probe = await probeLlamaCpp();
+        if (!probe.reachable) {
+          throw new ApiError(400, `Cannot enable llama.cpp routing: ${probe.reason ?? "server unreachable"}`);
+        }
+      }
+
+      let prev = false;
+      try {
+        const priorGlobal = await store.getGlobalSettingsStore().getSettings();
+        prev = priorGlobal.useLlamaCpp === true;
+      } catch {
+        // best effort
+      }
+
+      const settings = await store.updateGlobalSettings({ useLlamaCpp: enabled });
+      invalidateAllGlobalSettingsCaches();
+      const engineManager = options?.engineManager;
+      if (engineManager) {
+        for (const engine of engineManager.getAllEngines().values()) {
+          engine.getTaskStore().getGlobalSettingsStore().invalidateCache();
+        }
+      }
+
+      const next = settings.useLlamaCpp === true;
+      if (options?.onUseLlamaCppToggled && prev !== next) {
+        try {
+          options.onUseLlamaCppToggled(prev, next);
+        } catch (hookErr) {
+          console.warn(
+            `[auth/llama-cpp] onUseLlamaCppToggled callback threw: ${hookErr instanceof Error ? hookErr.message : String(hookErr)}`,
+          );
+        }
+      }
+
+      res.json({ enabled: next, restartRequired: false });
+    } catch (err: unknown) {
+      if (err instanceof ApiError) {
+        throw err;
+      }
+      rethrowAsApiError(err);
+    }
+  });
+
+  router.get("/providers/llama-cpp/status", async (_req, res) => {
+    try {
+      const probe = await probeLlamaCpp();
+      let enabled = false;
+      if (store) {
+        try {
+          const globalSettings = await store.getGlobalSettingsStore().getSettings();
+          enabled = globalSettings.useLlamaCpp === true;
+        } catch {
+          // best-effort
+        }
+      }
+      const extension = options?.getLlamaCppExtensionStatus?.() ?? null;
+      const ready = enabled && probe.reachable && (extension === null || extension.status === "ok");
+      res.json({
+        enabled,
+        extension,
+        ready,
+        server: {
+          available: probe.reachable,
+          url: probe.url,
+          hasApiKey: probe.hasApiKey,
+          reason: probe.reason,
+        },
+      });
+    } catch (err: unknown) {
+      if (err instanceof ApiError) {
+        throw err;
+      }
+      rethrowAsApiError(err);
+    }
+  });
+
   /**
    * POST /api/auth/login
    * Initiates OAuth login for a provider.

packages/dashboard/src/routes/register-model-routes.ts

@@ -13,6 +13,7 @@ export const registerModelRoutes: ApiRouteRegistrar = (ctx) => {
     let defaultModelId: string | undefined;
     let useClaudeCli = false;
     let useDroidCli = false;
+    let useLlamaCpp = false;
     let resolvedPlanningProvider: string | undefined;
     let resolvedPlanningModelId: string | undefined;
     if (store) {
@@ -25,6 +26,7 @@ export const registerModelRoutes: ApiRouteRegistrar = (ctx) => {
         defaultModelId = globalSettings.defaultModelId;
         useClaudeCli = globalSettings.useClaudeCli === true;
         useDroidCli = globalSettings.useDroidCli === true;
+        useLlamaCpp = globalSettings.useLlamaCpp === true;
 
         const mergedSettings = await store.getSettingsFast();
         const resolvedPlanningModel = resolvePlanningSettingsModel(mergedSettings);
@@ -82,6 +84,9 @@ export const registerModelRoutes: ApiRouteRegistrar = (ctx) => {
       if (!useDroidCli) {
         models = models.filter((m) => m.provider !== "droid-cli");
       }
+      if (!useLlamaCpp) {
+        models = models.filter((m) => m.provider !== "llama-server");
+      }
 
       res.json({
         models,

packages/dashboard/src/routes/register-settings-memory-routes.ts

@@ -1615,6 +1615,14 @@ export function registerSettingsMemoryRoutes(ctx: ApiRoutesContext, deps: Settin
         // Best-effort: on read failure assume false so a flip-on still fires.
       }
 
+      let prevUseLlamaCpp = false;
+      try {
+        const priorGlobal = await store.getGlobalSettingsStore().getSettings();
+        prevUseLlamaCpp = priorGlobal.useLlamaCpp === true;
+      } catch {
+        // Best-effort: on read failure assume false so a flip-on still fires.
+      }
+
       const settings = await store.updateGlobalSettings(req.body);
       // Invalidate global settings caches in all project-scoped stores so the
       // next GET /settings?projectId=xxx reads fresh values from disk rather
@@ -1653,6 +1661,17 @@ export function registerSettingsMemoryRoutes(ctx: ApiRoutesContext, deps: Settin
         }
       }
 
+      const nextUseLlamaCpp = settings.useLlamaCpp === true;
+      if (options?.onUseLlamaCppToggled && prevUseLlamaCpp !== nextUseLlamaCpp) {
+        try {
+          options.onUseLlamaCppToggled(prevUseLlamaCpp, nextUseLlamaCpp);
+        } catch (hookErr) {
+          runtimeLogger.warn(
+            `onUseLlamaCppToggled callback threw: ${hookErr instanceof Error ? hookErr.message : String(hookErr)}`,
+          );
+        }
+      }
+
       res.json(settings);
     } catch (err: unknown) {
       if (err instanceof ApiError) {

packages/dashboard/src/server.ts

@@ -287,6 +287,8 @@ export interface ServerOptions {
    * settings PUT to fail.
    */
   onUseDroidCliToggled?: (prev: boolean, next: boolean) => void;
+  /** Called when the user toggles the `useLlamaCpp` global setting. */
+  onUseLlamaCppToggled?: (prev: boolean, next: boolean) => void;
   /**
    * Returns the host's last-observed resolution of the bundled `droid-cli`
    * extension wiring. Populated by serve/daemon/dashboard startup checks.
@@ -305,6 +307,17 @@ export interface ServerOptions {
         reason?: string;
       }
     | null;
+  /** Returns the host's last-observed resolution of the bundled
+   * `@fusion/pi-llama-cpp` extension wiring. Populated by startup checks.
+   */
+  getLlamaCppExtensionStatus?: () =>
+    | {
+        status: "ok" | "not-installed" | "missing-entry" | "error";
+        path?: string;
+        packageVersion?: string;
+        reason?: string;
+      }
+    | null;
   /** Optional SkillsAdapter for skills discovery, execution toggling, and catalog fetching */
   skillsAdapter?: SkillsAdapter;
   /** Daemon mode configuration with bearer token authentication.

packages/pi-llama-cpp/index.ts

@@ -0,0 +1,38 @@
+import type { ExtensionAPI } from "@mariozechner/pi-coding-agent";
+import {
+  DEFAULT_CONTEXT_WINDOW,
+  DEFAULT_MAX_TOKENS,
+  PROVIDER_ID,
+  PROVIDER_NAME,
+} from "./src/constants.js";
+import { resolveLlamaServerApiKey, resolveLlamaServerUrl } from "./src/resolver.js";
+import { isLlamaServerReady, listLlamaModels } from "./src/retriever.js";
+
+export default async function (pi: ExtensionAPI): Promise<void> {
+  const cwd = process.cwd();
+  if (!(await isLlamaServerReady(cwd))) {
+    return;
+  }
+
+  const [url, models, apiKey] = await Promise.all([
+    resolveLlamaServerUrl(cwd),
+    listLlamaModels(cwd),
+    resolveLlamaServerApiKey(),
+  ]);
+
+  pi.registerProvider(PROVIDER_ID, {
+    name: PROVIDER_NAME,
+    baseUrl: `${url}/v1`,
+    api: "openai-completions",
+    apiKey: apiKey ?? "",
+    models: models.map((model) => ({
+      id: model.id,
+      name: model.id,
+      reasoning: true,
+      input: ["text", "image"] as Array<"text" | "image">,
+      cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 },
+      contextWindow: DEFAULT_CONTEXT_WINDOW,
+      maxTokens: DEFAULT_MAX_TOKENS,
+    })),
+  });
+}

packages/pi-llama-cpp/package.json

@@ -0,0 +1,36 @@
+{
+  "name": "@fusion/pi-llama-cpp",
+  "version": "0.17.2",
+  "description": "First-party Fusion pi extension for llama.cpp HTTP server integration.",
+  "license": "MIT",
+  "private": true,
+  "type": "module",
+  "main": "index.ts",
+  "keywords": [
+    "pi-package",
+    "fusion",
+    "llama-cpp"
+  ],
+  "pi": {
+    "extensions": [
+      "index.ts"
+    ]
+  },
+  "repository": {
+    "type": "git",
+    "url": "https://github.com/Runfusion/Fusion",
+    "directory": "packages/pi-llama-cpp"
+  },
+  "peerDependencies": {
+    "@mariozechner/pi-coding-agent": "*"
+  },
+  "devDependencies": {
+    "@types/node": "^22.0.0",
+    "typescript": "^5.7.0",
+    "vitest": "^3.0.0"
+  },
+  "scripts": {
+    "test": "vitest run --reporter=dot",
+    "typecheck": "tsc --noEmit"
+  }
+}

packages/pi-llama-cpp/src/__tests__/index.test.ts

@@ -0,0 +1,44 @@
+import { describe, expect, it, vi } from "vitest";
+import extension from "../../index.js";
+
+vi.mock("../retriever.js", () => ({
+  isLlamaServerReady: vi.fn(),
+  listLlamaModels: vi.fn(),
+}));
+vi.mock("../resolver.js", () => ({
+  resolveLlamaServerUrl: vi.fn(),
+  resolveLlamaServerApiKey: vi.fn(),
+}));
+
+import { isLlamaServerReady, listLlamaModels } from "../retriever.js";
+import { resolveLlamaServerApiKey, resolveLlamaServerUrl } from "../resolver.js";
+
+describe("pi-llama-cpp extension", () => {
+  it("does not register provider when server is offline", async () => {
+    vi.mocked(isLlamaServerReady).mockResolvedValue(false);
+    const registerProvider = vi.fn();
+
+    await extension({ registerProvider } as never);
+    expect(registerProvider).not.toHaveBeenCalled();
+  });
+
+  it("registers llama-server provider when server is reachable", async () => {
+    vi.mocked(isLlamaServerReady).mockResolvedValue(true);
+    vi.mocked(resolveLlamaServerUrl).mockResolvedValue("http://127.0.0.1:8080");
+    vi.mocked(resolveLlamaServerApiKey).mockResolvedValue("abc");
+    vi.mocked(listLlamaModels).mockResolvedValue([{ id: "qwen" }]);
+    const registerProvider = vi.fn();
+
+    await extension({ registerProvider } as never);
+
+    expect(registerProvider).toHaveBeenCalledTimes(1);
+    expect(registerProvider).toHaveBeenCalledWith(
+      "llama-server",
+      expect.objectContaining({
+        api: "openai-completions",
+        baseUrl: "http://127.0.0.1:8080/v1",
+        apiKey: "abc",
+      }),
+    );
+  });
+});

packages/pi-llama-cpp/src/__tests__/resolver.test.ts

@@ -0,0 +1,65 @@
+import { afterEach, beforeEach, describe, expect, it, vi } from "vitest";
+import {
+  resetLlamaResolverCache,
+  resolveLlamaServerApiKey,
+  resolveLlamaServerUrl,
+} from "../resolver.js";
+
+const readFileMock = vi.fn();
+vi.mock("node:fs/promises", () => ({
+  readFile: (...args: unknown[]) => readFileMock(...args),
+}));
+
+describe("resolveLlamaServerUrl", () => {
+  const originalEnv = process.env;
+
+  beforeEach(() => {
+    vi.resetAllMocks();
+    resetLlamaResolverCache();
+    process.env = { ...originalEnv };
+    delete process.env.LLAMA_SERVER_URL;
+  });
+
+  afterEach(() => {
+    process.env = originalEnv;
+  });
+
+  it("uses project config first", async () => {
+    readFileMock.mockResolvedValueOnce('{"url":"http://localhost:8081/"}');
+    const url = await resolveLlamaServerUrl("/tmp/project");
+    expect(url).toBe("http://localhost:8081");
+  });
+
+  it("falls back to env var", async () => {
+    readFileMock.mockRejectedValueOnce(new Error("missing"));
+    process.env.LLAMA_SERVER_URL = "http://localhost:9999/";
+    const url = await resolveLlamaServerUrl("/tmp/project");
+    expect(url).toBe("http://localhost:9999");
+  });
+
+  it("falls back to default", async () => {
+    readFileMock.mockRejectedValue(new Error("missing"));
+    const url = await resolveLlamaServerUrl("/tmp/project");
+    expect(url).toBe("http://127.0.0.1:8080");
+  });
+});
+
+describe("resolveLlamaServerApiKey", () => {
+  beforeEach(() => {
+    vi.resetAllMocks();
+  });
+
+  it("returns undefined when not configured", async () => {
+    readFileMock.mockRejectedValueOnce(new Error("missing"));
+    await expect(resolveLlamaServerApiKey()).resolves.toBeUndefined();
+  });
+
+  it("returns provider key for llama-server", async () => {
+    readFileMock.mockResolvedValueOnce(
+      JSON.stringify({
+        "llama-server": { type: "api_key", key: "secret-token" },
+      }),
+    );
+    await expect(resolveLlamaServerApiKey()).resolves.toBe("secret-token");
+  });
+});

packages/pi-llama-cpp/src/constants.ts

@@ -0,0 +1,5 @@
+export const PROVIDER_ID = "llama-server";
+export const PROVIDER_NAME = "Llama.cpp";
+export const DEFAULT_LLAMA_SERVER_URL = "http://127.0.0.1:8080";
+export const DEFAULT_MAX_TOKENS = 32000;
+export const DEFAULT_CONTEXT_WINDOW = 128000;

packages/pi-llama-cpp/src/resolver.ts

@@ -0,0 +1,62 @@
+import { readFile } from "node:fs/promises";
+import { join } from "node:path";
+import { DEFAULT_LLAMA_SERVER_URL, PROVIDER_ID } from "./constants.js";
+
+type AuthFile = Record<string, { type?: string; key?: string } | undefined>;
+
+let cachedUrl: string | null = null;
+
+async function readJson<T>(path: string): Promise<T | null> {
+  try {
+    const raw = await readFile(path, "utf-8");
+    return JSON.parse(raw) as T;
+  } catch {
+    return null;
+  }
+}
+
+function normalizeUrl(url: string): string {
+  return url.trim().replace(/\/+$/, "");
+}
+
+export async function resolveLlamaServerUrl(cwd: string): Promise<string> {
+  if (cachedUrl) return cachedUrl;
+
+  const projectCfg = await readJson<{ url?: string }>(
+    join(cwd, ".pi", "llama-server.json"),
+  );
+  if (projectCfg?.url) {
+    cachedUrl = normalizeUrl(projectCfg.url);
+    return cachedUrl;
+  }
+
+  const envUrl = process.env.LLAMA_SERVER_URL;
+  if (envUrl) {
+    cachedUrl = normalizeUrl(envUrl);
+    return cachedUrl;
+  }
+
+  const globalCfg = await readJson<{ llamaServerUrl?: string }>(
+    join(process.env.HOME ?? ".", ".pi", "agent", "settings.json"),
+  );
+  if (globalCfg?.llamaServerUrl) {
+    cachedUrl = normalizeUrl(globalCfg.llamaServerUrl);
+    return cachedUrl;
+  }
+
+  cachedUrl = DEFAULT_LLAMA_SERVER_URL;
+  return cachedUrl;
+}
+
+export async function resolveLlamaServerApiKey(): Promise<string | undefined> {
+  const authCfg = await readJson<AuthFile>(
+    join(process.env.HOME ?? ".", ".pi", "agent", "auth.json"),
+  );
+  const auth = authCfg?.[PROVIDER_ID];
+  const key = typeof auth?.key === "string" ? auth.key.trim() : "";
+  return key.length > 0 ? key : undefined;
+}
+
+export function resetLlamaResolverCache(): void {
+  cachedUrl = null;
+}

packages/pi-llama-cpp/src/retriever.ts

@@ -0,0 +1,43 @@
+import { resolveLlamaServerApiKey, resolveLlamaServerUrl } from "./resolver.js";
+
+export type LlamaModel = {
+  id: string;
+  object?: string;
+  owned_by?: string;
+};
+
+export type LlamaProviderModel = {
+  id: string;
+  name: string;
+  reasoning: boolean;
+  input: Array<"text" | "image">;
+  cost: { input: number; output: number; cacheRead: number; cacheWrite: number };
+  contextWindow: number;
+  maxTokens: number;
+};
+
+export async function llamaRpc<T>(endpoint: string, cwd = process.cwd()): Promise<T> {
+  const url = `${await resolveLlamaServerUrl(cwd)}${endpoint}`;
+  const apiKey = await resolveLlamaServerApiKey();
+  const response = await fetch(url, {
+    headers: apiKey ? { Authorization: `Bearer ${apiKey}` } : undefined,
+  });
+  if (!response.ok) {
+    throw new Error(`${response.status}: ${await response.text()}`);
+  }
+  return (await response.json()) as T;
+}
+
+export async function isLlamaServerReady(cwd = process.cwd()): Promise<boolean> {
+  try {
+    const status = await llamaRpc<{ status?: string }>("/health", cwd);
+    return status.status === "ok";
+  } catch {
+    return false;
+  }
+}
+
+export async function listLlamaModels(cwd = process.cwd()): Promise<LlamaModel[]> {
+  const response = await llamaRpc<{ data?: LlamaModel[]; models?: unknown }>("/models", cwd);
+  return Array.isArray(response.data) ? response.data : [];
+}

packages/pi-llama-cpp/tsconfig.json

@@ -0,0 +1,18 @@
+{
+  "compilerOptions": {
+    "target": "ESNext",
+    "module": "ESNext",
+    "moduleResolution": "bundler",
+    "strict": true,
+    "skipLibCheck": true,
+    "esModuleInterop": true,
+    "declaration": true,
+    "declarationMap": true,
+    "sourceMap": true,
+    "forceConsistentCasingInFileNames": true,
+    "resolveJsonModule": true,
+    "isolatedModules": true,
+    "noEmit": true
+  },
+  "include": ["src/**/*.ts", "index.ts"]
+}

packages/pi-llama-cpp/vitest.config.ts

@@ -0,0 +1,7 @@
+import { defineConfig } from "vitest/config";
+
+export default defineConfig({
+  test: {
+    globals: true,
+  },
+});

pnpm-lock.yaml

@@ -75,6 +75,9 @@ importers:
       '@fusion/pi-claude-cli':
         specifier: workspace:*
         version: link:../pi-claude-cli
+      '@fusion/pi-llama-cpp':
+        specifier: workspace:*
+        version: link:../pi-llama-cpp
       '@types/node':
         specifier: ^22.0.0
         version: 22.19.15
@@ -204,7 +207,7 @@ importers:
         version: link:../engine
       '@mariozechner/pi-coding-agent':
         specifier: ^0.72.1
-        version: 0.72.1(ws@8.20.0)(zod@3.25.76)
+        version: 0.72.1(@modelcontextprotocol/sdk@1.28.0(zod@3.25.76))(ws@8.20.0)(zod@3.25.76)
       '@types/multer':
         specifier: ^2.1.0
         version: 2.1.0
@@ -485,6 +488,22 @@ importers:
         specifier: ^3.0.0
         version: 3.2.4(@types/debug@4.1.13)(@types/node@22.19.15)(jiti@2.6.1)(jsdom@29.0.1)(tsx@4.21.0)(yaml@2.8.3)
 
+  packages/pi-llama-cpp:
+    dependencies:
+      '@mariozechner/pi-coding-agent':
+        specifier: '*'
+        version: 0.72.1(@modelcontextprotocol/sdk@1.28.0(zod@4.3.6))(ws@8.20.0)(zod@4.3.6)
+    devDependencies:
+      '@types/node':
+        specifier: ^22.0.0
+        version: 22.19.15
+      typescript:
+        specifier: ^5.7.0
+        version: 5.9.3
+      vitest:
+        specifier: ^3.0.0
+        version: 3.2.4(@types/debug@4.1.13)(@types/node@22.19.15)(jiti@2.6.1)(jsdom@29.0.1)(tsx@4.21.0)(yaml@2.8.3)
+
   packages/plugin-sdk:
     dependencies:
       '@fusion/core':
@@ -8108,6 +8127,19 @@ snapshots:
 
   '@exodus/bytes@1.15.0': {}
 
+  '@google/genai@1.46.0(@modelcontextprotocol/sdk@1.28.0(zod@3.25.76))':
+    dependencies:
+      google-auth-library: 10.6.2
+      p-retry: 4.6.2
+      protobufjs: 7.5.4
+      ws: 8.20.0
+    optionalDependencies:
+      '@modelcontextprotocol/sdk': 1.28.0(zod@3.25.76)
+    transitivePeerDependencies:
+      - bufferutil
+      - supports-color
+      - utf-8-validate
+
   '@google/genai@1.46.0(@modelcontextprotocol/sdk@1.28.0(zod@4.3.6))':
     dependencies:
       google-auth-library: 10.6.2
@@ -8526,9 +8558,9 @@ snapshots:
       - ws
       - zod
 
-  '@mariozechner/pi-agent-core@0.72.1(@modelcontextprotocol/sdk@1.28.0(zod@4.3.6))(ws@8.20.0)(zod@4.3.6)':
+  '@mariozechner/pi-agent-core@0.72.1(@modelcontextprotocol/sdk@1.28.0(zod@3.25.76))(ws@8.20.0)(zod@3.25.76)':
     dependencies:
-      '@mariozechner/pi-ai': 0.72.1(@modelcontextprotocol/sdk@1.28.0(zod@4.3.6))(ws@8.20.0)(zod@4.3.6)
+      '@mariozechner/pi-ai': 0.72.1(@modelcontextprotocol/sdk@1.28.0(zod@3.25.76))(ws@8.20.0)(zod@3.25.76)
       typebox: 1.1.32
     transitivePeerDependencies:
       - '@modelcontextprotocol/sdk'
@@ -8539,9 +8571,9 @@ snapshots:
       - ws
       - zod
 
-  '@mariozechner/pi-agent-core@0.72.1(ws@8.20.0)(zod@3.25.76)':
+  '@mariozechner/pi-agent-core@0.72.1(@modelcontextprotocol/sdk@1.28.0(zod@4.3.6))(ws@8.20.0)(zod@4.3.6)':
     dependencies:
-      '@mariozechner/pi-ai': 0.72.1(ws@8.20.0)(zod@3.25.76)
+      '@mariozechner/pi-ai': 0.72.1(@modelcontextprotocol/sdk@1.28.0(zod@4.3.6))(ws@8.20.0)(zod@4.3.6)
       typebox: 1.1.32
     transitivePeerDependencies:
       - '@modelcontextprotocol/sdk'
@@ -8598,6 +8630,28 @@ snapshots:
       - ws
       - zod
 
+  '@mariozechner/pi-ai@0.72.1(@modelcontextprotocol/sdk@1.28.0(zod@3.25.76))(ws@8.20.0)(zod@3.25.76)':
+    dependencies:
+      '@anthropic-ai/sdk': 0.91.1(zod@3.25.76)
+      '@aws-sdk/client-bedrock-runtime': 3.1036.0
+      '@google/genai': 1.46.0(@modelcontextprotocol/sdk@1.28.0(zod@3.25.76))
+      '@mistralai/mistralai': 2.2.1
+      chalk: 5.6.2
+      openai: 6.26.0(ws@8.20.0)(zod@3.25.76)
+      partial-json: 0.1.7
+      proxy-agent: 6.5.0
+      typebox: 1.1.32
+      undici: 7.24.6
+      zod-to-json-schema: 3.25.1(zod@3.25.76)
+    transitivePeerDependencies:
+      - '@modelcontextprotocol/sdk'
+      - aws-crt
+      - bufferutil
+      - supports-color
+      - utf-8-validate
+      - ws
+      - zod
+
   '@mariozechner/pi-ai@0.72.1(@modelcontextprotocol/sdk@1.28.0(zod@4.3.6))(ws@8.20.0)(zod@4.3.6)':
     dependencies:
       '@anthropic-ai/sdk': 0.91.1(zod@4.3.6)
@@ -8620,28 +8674,6 @@ snapshots:
       - ws
       - zod
 
-  '@mariozechner/pi-ai@0.72.1(ws@8.20.0)(zod@3.25.76)':
-    dependencies:
-      '@anthropic-ai/sdk': 0.91.1(zod@3.25.76)
-      '@aws-sdk/client-bedrock-runtime': 3.1036.0
-      '@google/genai': 1.46.0(@modelcontextprotocol/sdk@1.28.0(zod@4.3.6))
-      '@mistralai/mistralai': 2.2.1
-      chalk: 5.6.2
-      openai: 6.26.0(ws@8.20.0)(zod@3.25.76)
-      partial-json: 0.1.7
-      proxy-agent: 6.5.0
-      typebox: 1.1.32
-      undici: 7.24.6
-      zod-to-json-schema: 3.25.1(zod@3.25.76)
-    transitivePeerDependencies:
-      - '@modelcontextprotocol/sdk'
-      - aws-crt
-      - bufferutil
-      - supports-color
-      - utf-8-validate
-      - ws
-      - zod
-
   '@mariozechner/pi-coding-agent@0.62.0(@modelcontextprotocol/sdk@1.28.0(zod@4.3.6))(ws@8.20.0)(zod@4.3.6)':
     dependencies:
       '@mariozechner/jiti': 2.6.5
@@ -8708,11 +8740,11 @@ snapshots:
       - ws
       - zod
 
-  '@mariozechner/pi-coding-agent@0.72.1(@modelcontextprotocol/sdk@1.28.0(zod@4.3.6))(ws@8.20.0)(zod@4.3.6)':
+  '@mariozechner/pi-coding-agent@0.72.1(@modelcontextprotocol/sdk@1.28.0(zod@3.25.76))(ws@8.20.0)(zod@3.25.76)':
     dependencies:
       '@mariozechner/jiti': 2.6.5
-      '@mariozechner/pi-agent-core': 0.72.1(@modelcontextprotocol/sdk@1.28.0(zod@4.3.6))(ws@8.20.0)(zod@4.3.6)
+      '@mariozechner/pi-agent-core': 0.72.1(@modelcontextprotocol/sdk@1.28.0(zod@3.25.76))(ws@8.20.0)(zod@3.25.76)
-      '@mariozechner/pi-ai': 0.72.1(@modelcontextprotocol/sdk@1.28.0(zod@4.3.6))(ws@8.20.0)(zod@4.3.6)
+      '@mariozechner/pi-ai': 0.72.1(@modelcontextprotocol/sdk@1.28.0(zod@3.25.76))(ws@8.20.0)(zod@3.25.76)
       '@mariozechner/pi-tui': 0.72.1
       '@silvia-odwyer/photon-node': 0.3.4
       chalk: 5.6.2
@@ -8742,11 +8774,11 @@ snapshots:
       - ws
       - zod
 
-  '@mariozechner/pi-coding-agent@0.72.1(ws@8.20.0)(zod@3.25.76)':
+  '@mariozechner/pi-coding-agent@0.72.1(@modelcontextprotocol/sdk@1.28.0(zod@4.3.6))(ws@8.20.0)(zod@4.3.6)':
     dependencies:
       '@mariozechner/jiti': 2.6.5
-      '@mariozechner/pi-agent-core': 0.72.1(ws@8.20.0)(zod@3.25.76)
+      '@mariozechner/pi-agent-core': 0.72.1(@modelcontextprotocol/sdk@1.28.0(zod@4.3.6))(ws@8.20.0)(zod@4.3.6)
-      '@mariozechner/pi-ai': 0.72.1(ws@8.20.0)(zod@3.25.76)
+      '@mariozechner/pi-ai': 0.72.1(@modelcontextprotocol/sdk@1.28.0(zod@4.3.6))(ws@8.20.0)(zod@4.3.6)
       '@mariozechner/pi-tui': 0.72.1
       '@silvia-odwyer/photon-node': 0.3.4
       chalk: 5.6.2
@@ -8824,6 +8856,29 @@ snapshots:
       - bufferutil
       - utf-8-validate
 
+  '@modelcontextprotocol/sdk@1.28.0(zod@3.25.76)':
+    dependencies:
+      '@hono/node-server': 1.19.12(hono@4.12.9)
+      ajv: 8.18.0
+      ajv-formats: 3.0.1(ajv@8.18.0)
+      content-type: 1.0.5
+      cors: 2.8.6
+      cross-spawn: 7.0.6
+      eventsource: 3.0.7
+      eventsource-parser: 3.0.6
+      express: 5.2.1
+      express-rate-limit: 8.3.1(express@5.2.1)
+      hono: 4.12.9
+      jose: 6.2.2
+      json-schema-typed: 8.0.2
+      pkce-challenge: 5.0.1
+      raw-body: 3.0.2
+      zod: 3.25.76
+      zod-to-json-schema: 3.25.1(zod@3.25.76)
+    transitivePeerDependencies:
+      - supports-color
+    optional: true
+
   '@modelcontextprotocol/sdk@1.28.0(zod@4.3.6)':
     dependencies:
       '@hono/node-server': 1.19.12(hono@4.12.9)