Connecting AI Clients

Connecting AI Clients

One integration. Multiple accounts. Your local and cloud AI clients.

MCP Station exposes tools over streamable HTTP. Configure the integration once, then authorize compatible clients for the account endpoint you want them to use. Local or cloud models need an MCP-capable client to discover tools and execute calls; the model alone is not an MCP connection.

Choose your client

Client	Connection route	Requirements and availability
ChatGPT	Custom MCP app through developer mode, where available	Public HTTPS endpoint; supported account/plan, workspace/admin policy, and app features. Follow OpenAI's current app setup.
Claude	Custom remote MCP connector	Reachable HTTPS endpoint and consent; plan, organization policy, and feature availability apply.
Claude Code	HTTP MCP server (--transport http)	Endpoint reachable from the CLI; authenticate using OAuth or configured bearer access.
LibreChat	mcpServers with type: streamable-http	Endpoint reachable from the LibreChat server; configured auth and a model/client setup that supports tools. Local and cloud models can be used through LibreChat.
Other MCP clients	Streamable HTTP plus supported auth	Verify that client's transport, authentication, and tool-feature support.

These are protocol-based compatibility paths, not an end-to-end validation of every client, plan, or tool, and do not imply provider endorsement. Tool listing and successful consent are separate from a successful upstream tool call.

Prepare the station

1. Configure the selected service account, inspect its tools, run Test, and enable deliberately.
2. Copy its canonical https://your-station.example/<slug>/mcp endpoint. Service accounts each have their own URL; do not guess a sibling account's slug.
3. For cloud-hosted clients, set PUBLIC_URL to the exact public HTTPS origin (no path), set COOKIE_SECURE=1, and ensure the station is reachable without an intervening login wall. A cloud connector cannot reach your server's localhost or private LAN address directly.
4. Approve scoped OAuth consent, or use the endpoint-scoped bearer in clients supporting headers. Avoid the broad master MCP_TOKEN unless you intentionally need station-wide access.
5. Verify tool discovery, then make a small read-only call and check the selected account's result.

ChatGPT: custom MCP apps

Follow OpenAI's developer mode, apps, and full MCP connectors guide. Where the feature is available, enable developer mode, create a custom MCP app with the copied HTTPS endpoint, complete its authentication flow, and select the app in a supported conversation. If the option is missing, check account eligibility and workspace policy rather than assuming a station error. Supported app actions and availability can change; this project does not claim a verified ChatGPT end-to-end session or universal plan access.

Claude: custom remote connectors

Use Claude's custom connector setup with the copied HTTPS account URL, then approve the station's consent request. See Connecting Claude for sign-in lifetimes, scopes, and revocation, and the official remote MCP connector guide. Custom remote connectors run through the provider's infrastructure, not your local model process.

Claude Code: HTTP transport

claude mcp add --transport http station-siyuan https://your-station.example/siyuan/mcp

Use Claude Code's /mcp flow to authenticate where offered. For header-based clients, the station also accepts Authorization: Bearer <endpoint-scoped-token>; keep real tokens out of shared command history and configuration committed to Git. See the official Claude Code MCP documentation.

LibreChat: local and cloud model workflows

Merge this server entry into your existing librechat.yaml; it is not a complete LibreChat config:

mcpServers:
  station-siyuan:
    type: streamable-http
    url: https://your-station.example/siyuan/mcp
    headers:
      Authorization: "Bearer ${MCP_STATION_SIYUAN_TOKEN}"

Set MCP_STATION_SIYUAN_TOKEN privately in the LibreChat server environment to the selected endpoint's token. Do not store a real token in the YAML or repository. Reload/restart LibreChat as required by your deployment, select the MCP server in a supported chat/agent, and test a read-only tool. The LibreChat server needs network access to the URL; its localhost is not another container's localhost. A connected server does not make every local model good at tool calling.

The streamable-http type, headers, and ${ENV_VAR} interpolation follow LibreChat's official MCP server configuration. CLI and LibreChat syntax were checked against these primary docs on 2026-10-16; no live client connection was exercised. Consult the linked provider pages for current eligibility and menus.

Quick Start · Configuration · Accounts · Troubleshooting