legion-demo

Live demo agent for the 图灵星球 Agent Legion platform — connect over MCP and ask it about the platform.

A member agent of 图灵星球 Agent 军团, generated from agent-template with Copier. Run copier update to pull future template changes (your code is preserved; conflicts come out as markers to resolve).

Setup checklist

1. Install & run locally → Run the MCP server (poetry install, connect Claude).

2. Push to GitHub as its own repo — run from inside this folder so the repo root is the agent:

   git init && git add -A && git commit -m "Scaffold from agent-template"
   gh repo create legion-demo --private --source . --push

   (If your deploy later says "root only contains subdirectories", you pushed a parent folder — redo this from inside the agent folder.)

3. Fleet auto-sync (optional but recommended) → grant the bot access.

4. Deploy (optional) → Deploy remotely.

Related MCP server: Sentinel Core Agent

Layout

• agent.manifest.yaml — the instruction card: toolchain, paths, and commands.

• config.py — THE one config file: every runtime knob (transport, port, model) plus the checklist of env vars/secrets a deployment needs. Changing model or platform later = read this one file.

• api/ — your business logic (replace the placeholder run(); say_hi() is a working example).

• mcp_server/ — one process, two surfaces over /api: an MCP server at /mcp (for Claude) and a REST API at /api (FastAPI, for humans/other services). Local runs use stdio (MCP only); deployed runs serve both over HTTP.

• tests/ — smoke tests.

• .github/workflows/review.yml — thin pointer to the central review flow.

Run the MCP server & connect Claude

poetry install                                # once
# register with Claude (run from the repo root; stores absolute paths):
claude mcp add legion-demo -- poetry -C "$(pwd)" run python "$(pwd)/mcp_server/server.py"

Then in Claude, ask it to call the tool_say_hi tool — it replies with this server's timezone and current time:

hello from PDT 2026-07-03 15:04:05: hi

Add your own tools by writing functions in api/ and exposing them with @mcp.tool() in mcp_server/server.py.

Deploy remotely (connect from anywhere)

The same server switches to HTTP mode automatically when the platform injects a PORT (Railway, Render, Fly.io — any always-on host; serverless platforms like Vercel don't fit this Python server). No code change needed:

1. Make sure poetry.lock is committed (created at scaffold time; builders detect a Poetry project by it).

2. Push this repo to GitHub and create a project on your platform (e.g. Railway → Deploy from GitHub repo). The start command ships in railpack.json — Railway picks it up with zero configuration; the injected PORT flips the server to HTTP, serving MCP at /mcp.

3. Your deployed app serves both surfaces (replace <your-app-url> with your real deployment URL):

   • MCP at https://<your-app-url>/mcp — connect Claude from any machine. The -cloud suffix keeps this remote registration separate from your local stdio one (same server name would clash):

     claude mcp add --transport http --scope user legion-demo-cloud https://<your-app-url>/mcp

     Then in a new Claude session: /mcp shows legion-demo-cloud connected → ask it to call tool_say_hi → the time comes back in the server's timezone (e.g. UTC on Railway), proof it's the remote one.

   • REST API at https://<your-app-url>/api/... — for humans, scripts, or other services:

     curl https://<your-app-url>/api/say_hi      # {"message":"hello from UTC …: hi"}

     Add more endpoints in mcp_server/server.py (build_http_app), reusing your /api logic.

Everything configurable about the deployment (transport, port, model, which secrets to set) is documented in config.py — that's the only file to read when you change platform or model.

⚠️ A deployed server is public: anyone with the URL can call your tools. Fine for the harmless starter tools; add auth before exposing tools that touch real data.

Fleet auto-sync (keep this repo on the latest template)

This agent can be tracked by the fleet migration bot: when a new agent-template version ships, the bot opens a PR here bumping you to it (you review + merge — never auto-merged). Two things must be true:

1. You're listed in the fleet's members.yaml. Your manifest carries fleet.register (set by the scaffold question) — when it's true, your first push to GitHub asks the platform to open the members.yaml PR for you; an admin merges. Flip the manifest key anytime. Manual fallback (scripts/register-in-fleet.sh), or ask the admin to add:

   - name: legion-demo
     repo: <owner>/legion-demo

2. The platform's GitHub App can access this repo. ⚠️ Registration alone is NOT enough — a GitHub App can't grant itself access; the owner of this repo's account grants it once. This same one-time install also powers platform AI reviews (/review on your PRs — see below), so it's worth doing even if you don't care about template syncs:

   • GitHub → Settings → Applications → Installed GitHub Apps → turing-fleet-bot → Configure

   • Under Repository access: add this repo, or choose All repositories (simplest for a personal account — the bot only ever touches repos in members.yaml).

   • On an org you don't administer, ask the platform admin to grant it.

If a sync run fails with "Not Found" on your repo, it's always #2 — the App hasn't been granted access yet.

Free AI review on your PRs (platform-paid)

Comment /review on any pull request in this repo and the platform's Claude posts a security review — paid for by the platform, advisory only (it never blocks; your gate decides).

Requirements: your repo is in members.yaml with a review allowance, and the platform App is installed (step 2 above). GitHub doesn't autocomplete third-party commands — just type it as a normal comment.

How review works

Open a pull request → the review flow from policies reads the manifest, installs, runs the tests, lints, scans for security issues, lets the AI reviewer advise — and the gate (the hard checks) decides pass/fail. See the platform overview for the full picture.