fabric-runtime/pi-fabric

What it solves

Pi Fabric solves the problem of coordinating many different capabilities — Pi's core tools, MCP servers, extension tools, agents, actors, and workflows — into a single coherent execution. Instead of stitching together separate calls and managing state manually, you write one TypeScript or Python program that orchestrates everything, with built-in guardrails, parallelism, and a live dashboard.

How it works

You ask Pi in plain language, and Pi writes a single program (default TypeScript, optionally Python) that calls the required tools and agents. The program runs in a sandboxed kernel (QuickJS for TypeScript, Monty for Python) and is statically checked before execution. The program can call host providers for agents, actors, and durable coordination, and returns a result that goes back into the conversation. Independent calls run in parallel, and the whole system is governed by a unified schema validation and policy pipeline. You can also configure a CPython escape hatch for trusted native code, with explicit isolation requirements.

Who it’s for

This is for developers and teams using the Pi coding agent who need to run complex, multi-step agentic workflows — parallel audits, durable supervisors, decision councils, or resumable step graphs — without juggling many separate tools or writing glue code. It’s also for anyone who wants a single programmable runtime that composes MCP servers, model providers, and agent patterns under one API.

Highlights

• One executable tool (fabric_exec) composes core tools, MCP servers, and captured extension tools from a single program.
• Two sandboxed kernels: checked TypeScript (default) or Python (Monty), with an explicit CPython escape hatch.
• Built-in agent runtime: one-shot workers, durable resident agents, event-driven actors, councils, and recursive queries.
• Guardrails: approvals, isolation, timeouts, concurrency limits, recursion depth, and shared cost budgets.
• Native TUI with live activity panel, interactive dashboard, and settings.
• Achieved 100% on ARC-AGI-3 across all 25 environments in a 22.4-hour session.