MCP for coding agents
mixel mcp gives Claude Code, Cursor and other agents the whole build loop as small tools, so they check their own work.
mixel mcp is a local MCP server. It runs on your computer over stdio, needs no account, and reads and writes app folders through the same code as the mixel command. An agent that uses it gets error codes with one-line fixes instead of reading docs, which keeps it fast and its context small.
Add it to your agent
claude mcp add mixel -- mixel mcpWhat the agent can do
| Tool | Does |
|---|---|
start | The rules in short, and what to call next. Agents call it first. |
app_new | Creates an app folder that checks clean, with a test |
app_check | The checks Mixel and reviewers apply. fix: true adds missing permissions. |
app_describe | What Mixel will see: each action's input and the permissions |
app_run | Runs one action on real files, and returns the result and a log tail |
app_test | Checks, then runs tests/, and returns only the failures |
docs_search, docs_read | Finds and reads one section of these docs |
ui_components | The design-system components and tokens for a page |
app_status | Review state and reviewer notes, after mixel login |
library_overview, library_search, propose_crate | Your own Library, as your AI assistant sees it. Off unless you turn it on in Mixel. Needs mixel 0.3.5 or later. |
AGENTS.md, each docs page and each error code are also there as resources (mixel://agents.md, mixel://docs/{slug}, mixel://errors/{code}).
What it won't do
It never publishes. Sending an app for review stays your decision, with mixel publish in your terminal. mixel mcp --read-only offers only the tools that change nothing.
A good first prompt
Build a Mixel app in this folder that slices a song into equal pieces for a sampler. Use the mixel MCP tools: call start first, and keep calling app_check until it has no errors.
The 15-minute tutorial goes from there to a submitted app.