Connectors (remote MCP servers)

Connect remote MCP servers to the chat, and the model can use their tools: look up issues, query a database, read docs, whatever the server offers. You add a server by its URL; nothing runs on your computer.

MCP (Model Context Protocol) is the open standard many services use to expose tools to AI models.

Add a server

  1. In Chat, click Connectors.
  2. Enter:
    • Name: a short name, e.g. github (lowercase letters, digits, -, _). Tools show up as <name>__<tool>.
    • URL: the server's MCP endpoint, e.g. https://mcp.example.com/mcp.
    • Sign-in, one of:
      • API key or token: a header, usually Authorization with the value Bearer <your token>. Some servers use another header, such as X-API-Key.
      • Sign in with the service (OAuth): for servers that have you approve access in the browser. After Add, a window opens at the service; approve there and it closes by itself. The connector then shows ✓ Signed in.
      • None: for public servers.
  3. Click Test to connect and list its tools (for a token), then Add.

Turn on Web in the composer to use connectors. The toggle reads Web + N connectors while any are on. Each connector has its own on/off switch.

In a reply, each tool call shows as a chip, e.g. github · list_issues ✓. Hover it for the result. If a server can't be reached, the reply says so and the chat carries on without it.

Finding servers

Quick add in the Connectors dialog fills in a few well-known remote servers:

Server What it does Sign-in
DeepWiki Docs and Q&A for any public GitHub repo none
Context7 Up-to-date documentation for libraries none
Hugging Face Search models, datasets and Spaces none
Cloudflare Docs Cloudflare's documentation none
Notion, Linear, Sentry, Atlassian (Jira & Confluence) Work with your own workspace sign in with the service
GitHub Repos, issues, pull requests a personal access token
Stripe Your Stripe account a secret or restricted key

For more, browse a directory and pick servers marked remote or hosted (their address is an https://…/mcp or …/sse URL):

What works
Transports Streamable HTTP (current MCP) and the older HTTP + SSE transport (/sse endpoints). The gateway tries the current one first.
Auth A header (API key or personal access token), or OAuth sign-in following the MCP authorization spec: the gateway finds the service's sign-in endpoints, registers itself as a client, signs you in with PKCE, and refreshes the token before it expires. Services that need a client registered by hand (no dynamic client registration) aren't supported for sign-in; use a token for those.
Tools All of a server's tools. Text results go to the model; images in results are left out.
Limits Up to 10 connectors per user. A tool call times out after 2 minutes.

Stdio servers (the kind you run locally with npx … or uvx …) aren't connectors. Use a hosted version of the server if there is one.

Security

  • Your header values and sign-in tokens are secret. The gateway stores them server-side, encrypted, and never shows them again; the list shows only the last four characters of a header. They're sent only to that server, with each chat request that uses connectors. Sign out (or Remove) deletes the tokens; to revoke the gateway's access entirely, also remove it in the service's own settings.
  • If a sign-in expires and can't be refreshed, the chat says so; click Sign in again.
  • A connector's tools act with your token's permissions. The model decides when to call them, based on your conversation. Give connectors tokens with only the access they need (read-only where possible).
  • Tool results go into the conversation, and a server's text can try to steer the model. Add servers you trust.
  • Connectors must be on the public internet and use https://. Only admins can add servers on the local network or over plain http://.

For admins

Connectors run in the agent service (seed-agent) that powers the chat's Web mode. The gateway stores each user's servers with their account and passes the enabled ones to the agent with each request; the agent connects, lists the tools, and calls them during the conversation. The agent refuses connections to loopback, private and link-local addresses for non-admins, checked after DNS resolution.

API: GET /api/mcp lists the signed-in user's servers (header values masked); POST /api/mcp with {"action": "add" | "update" | "delete" | "test" | "sign_out", ...} manages them ("auth": "oauth" on add for sign-in). Sign-in runs through GET /api/mcp/oauth/start?id=<id> and returns to /api/mcp/oauth/callback, on the address the browser used; the service must allow that address as a redirect (dynamic registration registers it).

Sign in to LLM Gateway →