Connect to the remote MCP server

Team memory with zero local install. One hosted endpoint, the same two tools (memory_query / memory_write), any HTTP-capable MCP client.

The endpoint

https://threadctx.dev/mcp

Transport: Streamable HTTP. Auth: your team API key as a bearer token (Authorization: Bearer tctx_…) — the same key your team already uses for cloud mode. Requests without a valid key get a 401 and nothing else; memory is scoped to your team and never mixed across tenants.

Claude Code

claude mcp add --transport http threadctx https://threadctx.dev/mcp \
  --header "Authorization: Bearer tctx_YOUR_KEY"

Prefer working offline or per-repo? The npm package (npx threadctx-mcp) does the same job as a local process with a local-only mode — see the Claude Code guide.

Cursor

Add to .cursor/mcp.json (project) or ~/.cursor/mcp.json (global):

{
  "mcpServers": {
    "threadctx": {
      "url": "https://threadctx.dev/mcp",
      "headers": { "Authorization": "Bearer tctx_YOUR_KEY" }
    }
  }
}

Any other MCP client

Anything that speaks Streamable HTTP and can send a custom header works the same way: point it at the endpoint, add the bearer header. Windsurf, Zed, LangGraph (via langchain-mcp-adapters), Microsoft Agent Framework, and custom in-house agents all qualify — see the enterprise guide for custom-stack details.

claude.ai and Claude Desktop connectors

The claude.ai connector UI requires OAuth rather than bearer headers. OAuth support (and a connectors-directory listing) is in the works — until then, Claude Desktop users can run the local package via the desktop extension, and Claude Code users can use the header setup above today.

Getting a key

Keys are issued when a team signs up — see pricing. Your key identifies your whole team: any teammate (or agent) using it reads and writes the same shared memory. What one agent learns, everyone's agents remember.