Limoni — Claude Code MCP server: what it does, how to install
Limoni enables AI agents to inspect and drive Go terminal applications using widget semantic trees rather than raw screen characters.
Apache-2.0 ★ 196 updated 2026-10-29
Open on GitHubWhat Limoni does
Limoni exposes a dedicated MCP interface so models can navigate terminal programs built in Go. Traditional console automation breaks whenever layouts shift by a single character. This server bypasses raw character grids by exposing structured widget hierarchies directly to the model.
Under the hood, the server exposes eight automation actions, including tree, click, type_text, and wait_for. The control channel runs over an internal domain socket that only appears when building the Go binary with the limoni_debug tag. Connections are restricted to the local user account, keeping sensitive input masked from automated reading.
Launching the binary without the debug build tag prevents the automation socket from ever opening.
What it helps with
- run automated end-to-end checks on terminal interfaces
- submit console forms and confirm dialogs through an agent
- target interactive widgets by label instead of terminal coordinates
- verify screen updates without hardcoded sleep timers
Example prompts for the agent
- "Inspect the current widget tree and click the submit button"
- "Wait for the status indicator to update and report the message"
- "Type the deployment name into the active task field"
How to install Limoni
Install
The MCP registry has no launch command for it: see the repository README
Step by step, with commands for each platform: How to add an MCP server to Claude Code, Codex, Cursor, VS Code, Gemini CLI and more
Questions about Limoni
How do you install Limoni for Claude Code?
Compile your Go target application with the limoni_debug tag to generate the local socket. Once the program starts, link your MCP client using the limoni-mcp helper utility.
Why use a semantic tree instead of terminal grid scraping?
Grid scraping breaks whenever an element moves by a single character cell. Addressing components through a semantic tree keeps interactions stable regardless of layout adjustments.
Can the agent read sensitive form inputs?
Secret fields block automated reading by design. An agent can send text into the field, but the interface prevents it from inspecting the entered value.