🧵 pi-fabric

A programmable tool and agent runtime for Pi

One type-checked program for tools, MCP, agents, workflows, actors, mesh, councils, and recursion.

You keep talking to Pi the way you always do. Fabric gives the model one programmable tool — fabric_exec — that it uses to compose Pi's core tools, MCP servers, captured extension tools, child agents, persistent actors, and durable coordination into a single type-checked TypeScript program. The program runs in a QuickJS sandbox by default; an explicit unsafe Node-process executor is available for trusted workloads that exceed WASM32 memory. Only the final result comes back to the conversation. Branching, loops, fan-out, and data flow become code the model writes and type-checks — not a stack of separate tool calls you have to orchestrate.

Why Fabric?

How it works

1. You ask in plain language, as usual.
2. Pi writes one program that calls the tools, agents, and MCP servers it needs. The program is type-checked before it runs.
3. Only the result returns to your conversation. Intermediate work stays in the sandbox and surfaces in the activity panel and dashboard.

Under the hood, the model writes something like this — you don't:

const [manifest, sources] = await Promise.all([
  pi.read({ path: "package.json" }),
  pi.find({ pattern: "**/*.ts", path: "src" }),
]);
return {
  package: JSON.parse(manifest).name,
  sourceCount: sources.split("\n").filter(Boolean).length,
};

Independent calls run in parallel; only the returned object enters the model context. Known providers use concise direct calls such as mcp.fal_ai.get_model_schema(...), memory.recall(...), state.get(), schema.status(), and compact.status(); tools.call({ ref, args }) remains the fallback for refs discovered or computed at runtime.

Install

Requires Node.js 24+ and Pi 0.80.6+. Fabric also checks a detectable Pi host version at startup and warns when an older host may ignore continuation APIs such as actor triggerTurn.

pi install npm:pi-fabric

From GitHub:

pi install git:github.com/monotykamary/pi-fabric

From a local checkout:

pnpm install
pnpm build
pi install /absolute/path/to/pi-fabric

For one development run:

pi -e /absolute/path/to/pi-fabric

What you can ask for

Advanced patterns are user-invoked and are not advertised for automatic selection. Run /skill:fabric-guide when you want one recommendation, or invoke the exact /skill:<name> yourself. Describing an ordinary coding task keeps Pi on the core fabric-exec path.

The foundation is the fabric-exec reference skill: the model loads it before its first fabric_exec call and again when a call errors on argument shape.

The dashboard

Fabric adds a live activity surface to Pi, no extra extension required:

• A compact widget above the chat (like pi-supervisor) whose header follows the current phase while its rows show active/completed agents, active actors, and their recent nested tool or code-change activity.
• /fabric (or /fabric dashboard) — Activity and Topology views where the user-facing Pi session is always present as Main. Queue/steer Main, active children, actors, and observed mesh agents; inspect bounded Run/Project-mesh topologies, paged full agent/actor transcripts, topics, state, and routes.
• /fabric settings — mirrors Pi's /settings and writes changes to fabric.json.

See the interface & commands reference for every view, keybinding, and slash command.

Reference

• Configuration — fabric.json, code modes, tool capture, approvals, and budgets.
• Interface & commands — dashboard, settings, keybindings, slash commands, and headless runs.
• Agents, actors & mesh — subagents, the Claude runner, transports, steering, persistent actors, global templates, councils, recursive queries, and durable coordination.
• External providers — the versioned provider protocol for extensions.
• Architecture & security — the host bridge, sandboxing, tool-call robustness, and limitations.
• Skills — the core-first invocation policy and user-invoked advanced patterns.

Development

pnpm install
pnpm typecheck
pnpm test
pnpm build

The test suite covers configuration, schema validation, provider dispatch, registered-tool interception and execution, QuickJS isolation, Pi built-in invocation, subagents, fake Claude stream-JSON and model discovery, workflows, durable mesh state, actor mailboxes and subscriptions, and Pi/Claude actor restoration. Claude fixtures never make a billable request.

License

MIT