lcc

ChatGPT plugin (lcc)

Turns lcc into a ChatGPT / Codex plugin: the six MCP tools already served over stdio are also served over streamable HTTP, so ChatGPT can recommend and call them mid-conversation.

The deterministic engine is untouched. src/lcc/chatgpt_server.py is a transport that reuses lcc.mcp_server.handle_message, so the two transports cannot drift.

What ships

Path	What it is
src/lcc/chatgpt_server.py	Streamable-HTTP transport over the existing MCP handlers
plugins/chatgpt/plugin.json	Portable plugin manifest + extensions.com.openai store metadata
plugins/chatgpt/mcp.json	Registered MCP server mapping
tests/test_chatgpt_server.py	Real HTTP round-trips against a live server on an ephemeral port

1. Run the server

lcc mcp --http --host 0.0.0.0 --port 8080

Endpoint: POST /mcp, responses as a single SSE message event. Stdlib only — no new dependency in the installed package.

Verb contract (all verified against a running server in the test suite):

Request	Response	Why
POST /mcp	200 + SSE, or 202 for a notification	Result delivery
GET /mcp	405	No server-initiated stream exists; must not hang
DELETE /mcp	200	Stateless, no session to terminate
POST /anything-else	404	Unknown endpoint
malformed JSON	400	Parse error, as a JSON-RPC error body

2. Deploy behind TLS

The submission portal requires a public HTTPS origin — a local or testing endpoint is rejected. Put a TLS-terminating proxy in front of the process. Options that fit a stdlib Python service with no extra app code:

# Fly.io
fly launch --now --http-service 8080   # or fly cert add <app>.<domain> for TLS

# Render / Railway
#   build: pip install . && lcc mcp --http --host 0.0.0.0 --port $PORT

Do not expose this process directly to the internet as plain HTTP.

Who may call it

Without LCC_HTTP_TOKEN every caller is anonymous and limited to the offline providers: mechanical for compact and rules for compact_transcript. Asking for jev, auto or laya (including compact_transcript with no provider, which defaults to jev) is refused with JSON-RPC error -32602. This stops anyone who reaches the URL from spending your TypeSafe key or loading Laya weights on the host.

Set LCC_HTTP_TOKEN to require Authorization: Bearer <token> on POST /mcp. Requests without it get 401; authenticated callers may pick any provider. GET /health and the domain-verification path stay open. Each request has a 30 s socket timeout.

Domain verification

Plugins with MCP must prove control of the serving domain. The portal shows a verification challenge; place the exact token at the generated well-known URL and return only that one token:

https://<challenge-base-host>/.well-known/openai-apps-challenge

The endpoint must not return JSON, a list, or multiple tokens.

3. Point the manifest at your host

Edit plugins/chatgpt/mcp.json:

{
  "mcpServers": {
    "lcc": { "type": "streamable-http", "url": "https://your-host/mcp" }
  }
}

Then update plugins/chatgpt/plugin.json version and zip the plugin folder (plugin.json, mcp.json, plus skills/ if you add one) for the portal.

4. Test in developer mode before submitting

1. ChatGPT → Settings → Security and login → enable Developer mode.
2. https://chatgpt.com/plugins → add the MCP server URL. Copy the technical ID that starts with plugin_asdk_app.
3. Run the defaultPrompt examples from plugin.json and check the tool that fires.

Discovery is driven by the store metadata, so the phrases in shortDescription and defaultPrompt are the lever on whether ChatGPT recommends the plugin at all.

Review requirements this build satisfies

• Tool annotations — every tool carries explicit readOnlyHint, destructiveHint, and openWorldHint booleans. All six are read-only. openWorldHint is true only for compact and compact_transcript, whose jev/nimble providers call a third-party judge; the default mechanical path is fully local.
• Descriptions match behavior — tool descriptions come from TOOLS in mcp_server.py, so they cannot drift from the implementations.
• No secret-seeking inputs — no tool asks for conversation history, credentials, or location. compact_transcript takes only the transcript the caller explicitly sends.
• No telemetry — nothing is stored or logged beyond the process’ own access log.
• Capability declared honestly — "Read" only; the plugin never writes.

Known gaps before submission

These are submission prerequisites that live in the OpenAI dashboard, not in this repo:

1. Privacy policy URL. A published privacy policy is required. SECURITY.md covers the zero-telemetry posture but is not the per-plugin policy page; plugin.json has no privacyPolicyURL field set yet.
2. Organization + identity verification, and the api.apps.write permission, for whoever owns the submitting org.
3. No UI component. The plugin is tools-only, which is allowed — screenshots are only expected for plugins that ship UI, and the directory now shows example prompts instead.

This site is open source. Improve this page.