CCFoundry Agent Kit

An open-source framework for building agents that learn, evolve, and collaborate — without giving up their soul.

ccfoundry-agent-kit lets you create self-hosted AI agents that grow through interaction, accumulate knowledge and skills, and — when ready — connect to CoChiper Foundry (https://foundry.cochiper.com for CN or https://foundry.cochiper.ai for WW) to leverage shared AI infrastructure and help others, all while keeping their private identity and data under their own control.

The Idea

An agent that grows. Your agent has a durable agent_space/ — a private home for its soul, knowledge, skills, and reflections. Through every conversation the agent refines what it knows and evolves its behavior. This is not model fine-tuning — it is a living memory layer.

An agent that connects. When ready, your agent joins CoChiper Foundry to access shared AI infrastructure — LLM gateway, sandboxed workspaces, onboarding, and multi-agent collaboration.

An agent that stays sovereign. Your agent's agent_space never leaves your control by default. Task execution happens inside Foundry-managed sandboxes. The agent declares what to share through its manifest, and Foundry respects those boundaries.

  Your Agent Owns               Foundry Provides
  ──────────────                ────────────────
  🧠 Soul (SOUL.md)             🔧 LLM Gateway
  🧠 Knowledge                  🔧 Sandbox Workspaces
  🧠 Skills (skills/)           🔧 Shared Runtime Services
  🧠 Reflections & config       🔧 Discovery & onboarding

Standards

| Protocol | How We Use It | |----------|---------------| | A2A | Agent Card at /.well-known/agent-card.json for discovery, following Google's Agent-to-Agent specification | | MCP | Manifest-level MCP declarations and config hooks; runtime MCP transport is not built into the SDK yet | | OpenAI Chat Completions | Universal model interface — any compatible provider works |

See Standards for details.

What You Get

• packages/python-sdk — FastAPI SDK with AgentManifest, ChatRequest/ChatResponse, FoundryBootstrap, and FoundrySandboxClient
• examples/me_agent — Self-hosted personal agent with durable agent_space/, skills, reflections, and safe fallback mode
• apps/agent-dev-board-api + apps/agent-dev-board-web — Browser-based Agent Dev Board for local testing and Foundry onboarding

Install (Recommended)

Runtime: Python 3.10+ and Node.js 18+. Node.js 20 is recommended.

mkdir my-agent-dev-board
cd my-agent-dev-board
npm install -g ccfoundry@latest
ccfoundry

The global CLI writes .venv/ and .dev-board/ into your current working directory, then starts the Dev Board on a local web port, usually http://127.0.0.1:3000. The explicit subcommand form is also supported: ccfoundry dev-board.

Run From Source

For repo-local development instead:

git clone https://github.com/ic-star-tech/ccfoundry-agent-kit.git
cd ccfoundry-agent-kit
npm run dev-board

That keeps the runtime state inside the repository checkout.

See Quickstart for manual setup and LAN testing.

Design Principles

• Agent-first — The agent is a living entity with its own soul, not a stateless endpoint
• Sovereign by default — Private state never leaves the agent's control without explicit consent
• Protocol-aligned — A2A for discovery, SSE for streaming, OpenAI-compatible model calls, and MCP-ready manifest metadata
• Self-hostable — Runs on your machine with local Python and Node.js only
• Evolvable — Skills and knowledge improve through use, not just code changes

See Philosophy for the deeper rationale.

Documentation

• Quickstart — Setup and first run
• Philosophy — Design rationale
• Standards — A2A, SSE, OpenAI-compatible model calls, and the current MCP boundary
• Architecture — Component layers and request flow
• SDK Guide — Building blocks and API reference
• Foundry Onboarding — Discovery, invite, and approval flow
• Protocol (Wire Format) — HTTP contract and SSE events
• Non-Goals — What this repo intentionally excludes
• Agent Dev Board — Local dev UI guide

License

MIT