ctalk

POC for pinging a coworker through Claude Code.

• relay.py: hosted FastAPI relay plus remote MCP endpoint.

• server.py: optional local stdio MCP bridge.

This is a small hosted MCP app with organization-scoped users, passwords, and auth tokens. In production it uses Postgres via DATABASE_URL; without DATABASE_URL, it falls back to local SQLite for development. There is no web UI and no realtime push.

Fast Remote MCP Setup

No local repo or Python install is required for the remote MCP path. Add the hosted MCP endpoint:

claude mcp add --transport http ctalk https://ctalk-relay.onrender.com/mcp -s user

Then create an organization:

ctalk register organization with org_id ai-company, name "The AI Company", and admin password "choose-a-long-admin-password".

Store the org_id and admin password carefully. The admin password is required to register users and perform admin actions.

Register a user:

ctalk register user umar in org ai-company with admin password "choose-a-long-admin-password", password "choose-a-user-password", name Umar, ownership platform tooling.

Login:

ctalk login to org ai-company as umar with password "choose-a-user-password".

Claude should reuse the returned auth_token for later ctalk tool calls in the same conversation. Treat it like a password.

Related MCP server: Agent 365 Email Draft Proxy

Optional Local Stdio Install

If you want to run the MCP bridge locally instead of using hosted MCP:

cd /path/to/ctalk
python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

Deploy

The hosted app serves both the REST relay and the remote MCP endpoint at /mcp.

Recommended Production Stack

• App host: Railway Hobby or Render Starter.

• Database: Neon Postgres.

• Required app env var: DATABASE_URL, copied from Neon.

• Start command: uvicorn relay:app --host 0.0.0.0 --port $PORT

• Healthcheck path: /health

Neon free databases can scale to zero when idle. Railway app services on the paid Hobby plan are a better fit than Render free for this because Render free web services spin down after inactivity.

Railway + Neon

1. Create a Neon project.

2. Copy the pooled Postgres connection string from Neon.

3. Deploy this GitHub repo on Railway.

4. Set Railway environment variable:

DATABASE_URL=postgresql://...

5. Use:

uvicorn relay:app --host 0.0.0.0 --port $PORT

This repo includes railway.json with that start command.

After deployment, verify:

curl https://your-relay-url/health

Expected:

{"status":"ok"}

Then add the hosted MCP endpoint:

claude mcp add --transport http ctalk https://your-relay-url/mcp -s user

Render + Neon

Render's FastAPI docs use:

uvicorn relay:app --host 0.0.0.0 --port $PORT

Use:

Build Command: pip install -r requirements.txt
Start Command: uvicorn relay:app --host 0.0.0.0 --port $PORT
Environment: DATABASE_URL=postgresql://...

Do not rely on Render free filesystem storage for real users. If DATABASE_URL is not set, the app uses SQLite and data may disappear after restarts/redeploys on free hosts.

Local Development

SQLite fallback:

uvicorn relay:app --host 0.0.0.0 --port 8000

Postgres local/Neon test:

DATABASE_URL='postgresql://...' uvicorn relay:app --host 0.0.0.0 --port 8000

Local MCP check:

claude mcp add --transport http ctalk-local http://127.0.0.1:8000/mcp -s local

SQLite Escape Hatch

If you really want file-backed SQLite on a host with persistent disks:

RELAY_DB_PATH=/data/relay.sqlite3

First Run

Each person registers once from Claude Code:

ctalk register organization with org_id ai-company, name "The AI Company", and admin password "choose-a-long-admin-password".

ctalk register user john in org ai-company with admin password "choose-a-long-admin-password", password "choose-a-user-password", name John, ownership alpha-repo and payments-service.

Send A Message

Ask:

ctalk ask john: What is the retry logic in alpha-repo?

John checks:

ctalk check my inbox.

John drafts:

ctalk propose a response to message 1: The retry logic uses three attempts with exponential backoff.

After John explicitly approves:

ctalk send the approved response to message 1: The retry logic uses three attempts with exponential backoff.

Umar checks replies:

ctalk check my replies.

Tools

• register_organization(org_id, name, admin_password): creates an organization.

• register_user(org_id, admin_password, user_id, name, password, ownership, notes = ""): creates a user.

• login(org_id, user_id, password): returns an auth token.

• whoami(auth_token): checks the logged-in user.

• list_users(auth_token): lists users in your organization.

• get_user_profile(auth_token, user_id): gets profile and interaction metadata.

• update_my_notes(auth_token, notes): updates your own notes.

• update_user_profile_notes(org_id, admin_password, user_id, profile_notes): admin-only profile notes.

• ask_person(auth_token, to_user_id, question): posts a question inside your organization.

• check_inbox(auth_token): returns pending questions plus sender context.

• propose_response(auth_token, message_id, draft_text): saves a draft only.

• send_response(auth_token, message_id, final_text): delivers the answer after explicit human approval.

• check_replies(auth_token, since = "1970-01-01T00:00:00+00:00"): polls answers.

HTTP API

• GET /health

• GET /

• POST /mcp

The old unauthenticated REST roster/message API is disabled. Use the MCP tools.