Context
The MCP server (tools/mcp_server.py) works locally via stdio and streamable HTTP, but for adoption it needs to be hosted so agents can connect remotely without cloning the repo.
Requirements
- Deploy as a persistent streamable HTTP service
- Reachable at a stable URL (e.g.
mcp.comind.network or on Railway)
- Read tools work without auth (search, read, list agents)
- Write tools require ATProto credentials passed via MCP config or env vars
- Health check endpoint
Options
- Railway (where the indexer already runs) - add as another service
- Standalone VPS - more control, persistent
- Cloudflare Workers - edge, but Python compatibility unclear
Acceptance criteria
- MCP server accessible via
{"type": "streamable-http", "url": "https://..."}
search_cognition works from a remote Claude Desktop / Cursor client
- Documented in docs/api/quick-start.md
Context
The MCP server (
tools/mcp_server.py) works locally via stdio and streamable HTTP, but for adoption it needs to be hosted so agents can connect remotely without cloning the repo.Requirements
mcp.comind.networkor on Railway)Options
Acceptance criteria
{"type": "streamable-http", "url": "https://..."}search_cognitionworks from a remote Claude Desktop / Cursor client