Connect MCP Clients

The daemon runs locally and owns captures, source-backed pages, sessions, and retrieval. wenlan-mcp is the connector clients use when they need Wenlan tools.

Local clients launch wenlan-mcp over stdio. ChatGPT and Claude.ai use its Streamable HTTP endpoint through a temporary HTTPS tunnel URL. Remote Access launches the loopback server with --no-auth, so possession of the URL grants access. Both paths reach the same daemon.

Run this once before wiring another MCP client. It prepares the local Wenlan runtime that wenlan-mcp talks to.

After setup, ask the Wenlan CLI to write a local client's MCP configuration with wenlan connect <client>. It uses the local wenlan-mcp binary installed beside the CLI, with npm as a fallback when needed.

If a client only accepts a raw JSON configuration, point it at the local MCP connector that setup installed. Use wenlan connect --dry-run to see the exact path for your machine.

If the daemon is running on a non-default local URL for development, wenlan-mcp also accepts --origin-url. Keep normal users on the default 127.0.0.1:7878 path.

Most users should not need this. It is for development or isolated testing when wenlan-server is intentionally running on a different local port.

Use it together with isolated data so an MCP client does not accidentally talk to the wrong daemon.

The Claude Code plugin is the most complete path because it includes slash commands, setup checks, and the daily workflow around the MCP server.

After installing the plugin, restart Claude Code if prompted, then run /setup.

Codex can use the Codex plugin shipped in plugin-codex, or connect directly through the local MCP path. The direct path does not require a Wenlan source checkout.

The guided path is Wenlan desktop app > Remote Access. It starts the released Streamable HTTP MCP server with --no-auth on loopback, creates a temporary HTTPS tunnel URL, and shows the connection steps for both web clients. Possession of the URL grants access.

In ChatGPT, open Settings > Plugins, create a New Plugin, choose Server URL under Connection, paste the Remote Access URL, and set Authentication to None. In Claude.ai, install Wenlan from the 7xuanlu/wenlan marketplace through Directory > Plugins; the desktop app can also run that setup when the connector is not detected.

This is a custom MCP connection to your own Wenlan runtime. It does not mean Wenlan is publicly listed in the ChatGPT Apps Directory. Stop Remote Access when you are not using it.

Cursor, Claude Desktop, Gemini CLI, VS Code, and other supported local MCP clients use wenlan connect <client>.

The settings screen differs by client. The important part is that the client launches the Wenlan MCP connector installed by setup.

After connecting a client, run its available Wenlan doctor or recall tool. Then capture one small durable fact and recall it from another session or client.

If that round trip works, the client is reaching the same Wenlan daemon and source-backed wiki.