corpus-agent-kit

Live demo: corpuslaw.us/autopilot — the Formation Autopilot chat, backed by this repo's agent running on Alibaba Cloud Function Compute (us-west-1, config in autopilot/deploy/alibaba/s.yaml). API health: curl https://formatiutopilot-bfmjghskwt.us-west-1.fcapp.run/healthz → ok. (The raw fcapp.run URL force-downloads HTML on FC's default domain — use the hosted page in a browser.)

Open-source connector kit for the Corpus legal platform:

• autopilot/ — Formation Autopilot: a Qwen Cloud agent that turns an ambiguous founder description into a cited launch checklist and a prefilled, human-approved formation handoff. (Qwen Hackathon Track 4 entry.)
• mcp-server/ — zero-dependency stdio MCP server exposing Corpus law search to Claude Desktop, Cursor, and any MCP client.
• widget/ — embeddable law-search widget (Preact, ~12 kB gzipped).

All three are thin clients of the hosted Corpus platform. The law corpus, hybrid search engine, human approval gate, and filing execution live in the hosted service — this repo never touches money or files anything.

Architecture

See docs/architecture.md for the Mermaid source.

What's open vs. what's hosted

Layer	This repo (Apache-2.0)	Hosted Corpus platform (closed)
Formation Autopilot agent	✅ autopilot/	—
Two-lane Qwen routing	✅ autopilot/src/	—
MCP stdio bridge (corpus-mcp)	✅ mcp-server/	—
Embeddable widget	✅ widget/	—
Law database (~186 K nodes, Aurora)	—	✅
Hybrid search engine + /api/mcp	—	✅
/formation checkout + GATE 2	—	✅
Stripe payment + state filing execution	—	✅

Two-lane model routing

The autopilot uses two Qwen models with different cost/quality profiles:

• Fast lane (QWEN_MODEL_FAST, default qwen-flash): all standard turns — intent parsing, law search, checklist generation. Low latency, low spend.
• Critical lane (QWEN_MODEL_CRITICAL, default qwen3.7-max): the final pre-handoff draft review only. Higher quality for the one turn that shapes the prefilled filing URL.

The lane switch is automatic. Routing logic lives in autopilot/src/agent.ts. The hourly spend cap (SPEND_CAP_TURNS_PER_HOUR) applies across both lanes.

Human in the loop (GATE 2)

The autopilot produces a prefilled draft URL — it never initiates a filing or charges a card. On the Corpus platform side, every order passes GATE 2: a human reviews and approves a snapshot of the exact filing payload, and that approval is cryptographically bound to a hash of the payload. If the payload changes by a single byte after approval, the gate rejects it. The agent is architecturally incapable of bypassing this.

Quickstart

Prerequisites

node >=20
npm >=9

Clone and install all workspaces:

git clone https://github.com/teakesdev/corpus-agent-kit.git
cd corpus-agent-kit
npm install

Environment variables

Copy and fill .env.example (required for autopilot; mcp-server and widget read CORPUS_BASE_URL / CORPUS_API_KEY from env at runtime):

cp .env.example .env
# edit .env — at minimum set QWEN_API_KEY and CORPUS_API_KEY

Variable	Default	Purpose
QWEN_API_KEY	(required)	Qwen Cloud API key
QWEN_BASE_URL	(required)	Qwen OpenAI-compatible base URL (e.g. https://dashscope-intl.aliyuncs.com/compatible-mode/v1)
QWEN_MODEL_FAST	qwen-flash	Fast-lane model — standard turns
QWEN_MODEL_CRITICAL	qwen3.7-max	Critical-lane model — pre-handoff review
CORPUS_BASE_URL	https://corpuslaw.us	Hosted Corpus platform base URL
CORPUS_API_KEY	(optional)	Corpus platform API key (optional — anonymous is rate-limited)
SPEND_CAP_TURNS_PER_HOUR	120	Abuse guard: max agent turns per hour
PORT	9000	HTTP server port (autopilot backend)

autopilot — Formation Autopilot agent

cp .env.example .env        # fill QWEN_API_KEY + CORPUS_API_KEY
npm install
npm run build               # compiles all workspaces
cd autopilot && npm start   # starts the autopilot HTTP server
# open http://localhost:9000

The server exposes:

• GET /healthz — liveness probe
• POST /api/chat — agent turn endpoint
• GET / — chat UI

Deploy to Alibaba Cloud Function Compute: see autopilot/deploy/alibaba/README.md.

mcp-server — stdio MCP bridge

Build locally:

cd mcp-server && npm install && npm run build

Add to your MCP client config (Claude Desktop, Cursor, etc.):

{
  "mcpServers": {
    "corpus-law": {
      "command": "node",
      "args": ["/absolute/path/to/corpus-agent-kit/mcp-server/dist/index.js"]
    }
  }
}

Replace /absolute/path/to/corpus-agent-kit with the actual path to your clone.

Note: The package name @corpus-agent-kit/mcp-server is reserved for a future npm publish; until then, build locally and point your MCP client at mcp-server/dist/index.js. HTTP MCP clients: the hosted endpoint https://corpuslaw.us/api/mcp supports the streamable-HTTP MCP transport directly — no local bridge needed.

This exposes three tools: search_law, get_law_node, list_coverage.

widget — embeddable law-search

cd widget
npm run build               # outputs dist/widget.js (loader) and dist/widget-app.js (app)

Embed in any page using the hosted loader. Point src at the Corpus host (a relative path would 404 on the host page's own origin) and pass your widget key (pk_…):

<script async src="https://corpuslaw.us/widget/widget.js"
        data-corpus-key="pk_live_…"
        data-corpus-origin="https://corpuslaw.us"></script>

Deploy to Alibaba Cloud

See autopilot/deploy/alibaba/README.md for Function Compute deploy instructions (Serverless Devs s CLI).

License

Apache-2.0 — see LICENSE.