filepath

Orchestration layer for AI agents. Liquid agents, liquid models, multiple interfaces. BYOK (bring your own key).

filepath is infrastructure for coordinating AI agents. Agents are interchangeable. Models are interchangeable. We provide the orchestration layer — Dashboard for humans, API/MCP/TypeScript for programs.

The Core Idea

Liquid Agents — Swap Claude Code for Codex for Shelley without changing your workflow. Same protocol, seamless handoff.

Liquid Models — Switch models mid-session. Start with Claude Sonnet, escalate to o3 for reasoning, fallback to Gemini for speed.

Orchestration Infrastructure — One platform, multiple interfaces:

• Dashboard — Visual tree, rich chat, human-in-the-loop
• REST API — Spawn agents, send messages, receive events
• WebSocket — Real-time streaming for live collaboration
• MCP + TypeScript — Agents calling agents programmatically

Quick Start

1. Sign up at https://myfilepath.com
2. Add your API key in Settings → Provider API Keys
   • Get an OpenRouter key at https://openrouter.ai/keys
   • Or use OpenAI, Anthropic, or any OpenAI-compatible provider
3. Create a session from the dashboard
4. Spawn an agent — pick any agent type, pick any model
5. Send a message — that's the agent's task
6. Watch it work — tool calls, file writes, commits, all in real-time chat

How It Works

1. Create a session (like a project folder). Optional git repo to clone.
2. Spawn an agent — pick its type (Claude Code, Cursor, Codex, Shelley, or BYO), pick its model (Claude, GPT, Gemini, 100+ more)
3. The agent works in an isolated container with your repo. Send it messages via chat.
4. Watch through rich chat — tool calls, file writes, commits, all inline
5. Agents spawn child agents. The tree grows. Click any node, see its conversation.
6. Close your laptop, open your phone. Same state, real-time sync.

BYOK Model: You bring your own LLM API keys. We don't charge for usage — your keys, your spend.

Stack

SvelteKit (Svelte 5) + Cloudflare Workers + Agents SDK + D1 + CF Sandbox + Alchemy

The filepath Agent Protocol (FAP)

Agents communicate via NDJSON over stdout/stdin. One JSON object per line, validated with Zod schemas.

This protocol makes agents liquid — any agent that speaks FAP runs on filepath, regardless of what's inside the container.

Container receives: repo at /workspace, task via FILEPATH_TASK env var, user messages on stdin.

Container emits: structured events to stdout — text, tool calls, commands, commits, spawn requests, status updates, handoffs.

Built-in agents (Shelley, Claude Code, Cursor, Codex, Amp) are reference implementations. BYO = bring a Dockerfile that speaks the protocol.

See NORTHSTAR.md for the full protocol spec.

Development

bun install
bun run dev          # localhost:5173
bun run deploy       # Alchemy (never wrangler)
bun run prd          # Run gates

Agent Memory (Deja)

This project uses deja for persistent memory across agent sessions.

Set DEJA_API_KEY in your .env (ask project owner for the key).

# Query memories at session start
curl -s -X POST https://deja.coey.dev/inject \
  -H "Authorization: Bearer $DEJA_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"context": "filepath current status", "format": "prompt", "limit": 7}'

# Store learnings after milestones
curl -s -X POST https://deja.coey.dev/learn \
  -H "Authorization: Bearer $DEJA_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"trigger": "when to surface", "learning": "what was learned", "confidence": 0.9}'

Documentation

• NORTHSTAR.md — The plan. Protocol spec, schema, UI components, execution phases.
• AGENTS.md — Agent catalog, architecture, development rules.
• docs/API-REFERENCE.md — API endpoints
• llms.txt — Context for AI tools

License

See LICENSE.