wenlan-mcp

MCP server for Wenlan. It lets Claude Code, Cursor, Codex, Claude Desktop, Gemini CLI, and other MCP clients read and write to the local Wenlan daemon through the Model Context Protocol.

Wenlan owns storage, search, embeddings, pages, and distill cycles. wenlan-mcp is the connector.

Install

Most users should install the runtime through the root README (npx -y wenlan setup on macOS Apple Silicon, the automated shell setup on Linux, or the matching release archive on Windows). Then use the product CLI to configure supported clients:

wenlan connect codex              # or: claude-code, cursor, claude-desktop, vscode, gemini
wenlan connect cursor --dry-run   # preview before editing JSON config

MCP-only setup gives agents tools for capture, recall, context, doctor, and page distillation. It does not install Claude Code slash skills like /brief, /handoff, /distill, or /setup; use the Wenlan plugin for that workflow.

If you only need the raw MCP connector config, add this to your MCP client:

{
  "mcpServers": {
    "wenlan": {
      "command": "npx",
      "args": ["-y", "wenlan-mcp"]
    }
  }
}

The npm wrapper auto-detects the host platform and downloads the matching prebuilt binary from the Wenlan release. Supported: macOS (arm64), Linux (x64, arm64; glibc), Windows (x64). Other targets require building the connector from source via cargo install wenlan-mcp; macOS Intel does not currently have a supported complete local runtime.

Or install a binary directly:

brew install 7xuanlu/tap/wenlan-mcp
cargo install wenlan-mcp

Then use:

{
  "mcpServers": {
    "wenlan": {
      "command": "wenlan-mcp"
    }
  }
}

wenlan-mcp expects the Wenlan daemon at http://127.0.0.1:7878 by default. Override it with:

wenlan-mcp --origin-url http://127.0.0.1:7879

Tools

Tool	Purpose
`context`	Load session context. Use at session start or major topic shifts.
`capture`	Save one durable memory: decision, lesson, gotcha, preference, fact, correction, or project context.
`recall`	Search memories and pages by natural-language query.
`distill`	Trigger page distillation for new clusters or a specific `page_id`.
`list_pending`	List unconfirmed memories waiting for review.
`confirm_memory`	Confirm a pending memory by `source_id`.
`forget`	Delete a memory by ID. Destructive.
`doctor`	Diagnose daemon reachability, setup mode, API key state, and on-device model state.

doctor mirrors wenlan doctor. It is diagnostic only and is not part of the memory loop.

Setup Modes

Wenlan works immediately in local memory mode: storage, search, recall, and MCP memory are available without a local model or API key.

Users can opt into more expensive distill cycles:

On-device model: private extraction and distillation after wenlan models install.
Anthropic key: richer extraction and page synthesis after wenlan keys set anthropic.

Agent Guidance

The MCP server ships tool instructions that tell agents to capture durable state proactively:

One idea per capture.
Include the why, not just the what.
Name people, projects, and tools explicitly.
Omit memory_type unless the agent is certain.
Do not store tool output, command logs, filler, or transient task state.

See src/tools.rs for the full instructions.

License

Apache-2.0.