atomicreps MCPサーバー

atomicreps

One short question about the thing you just built. Works inside Claude Code, GitHub Copilot, Cursor, Windsurf, Codex, or any MCP client.

You describe a bug, your coding agent changes the code, and Atomic Reps asks you one multiple-choice question about that change. You answer with a letter, from memory, and the server grades it.

The two-minute demo: four bugs fixed in Claude Code, one question after each.

https://github.com/user-attachments/assets/521c1bb3-9f75-4c25-8b71-ed763de7debc

Quick start

Requires Node 24+ and a free account. On an older Node, every command prints the version it needs instead of starting.

npx atomicreps

The first run is a setup wizard:

1. Pick what to practise, how often, and how hard.
2. Sign in: the terminal prints a code, and you type it at atomicreps.com/connect.
3. Pick your editors. connect writes only the entries you tick.

Then go back to work. When your agent finishes a task that changed files, the question appears after its answer:

Type the letter. The verdict comes back with a one-line explanation, and the next line names up to three other topics the session touched:

Every run after the first opens the terminal screen: your streak, this session, your topics, mutes and rate.

Commands

npx atomicreps            the terminal screen; the setup wizard on the first run
npx atomicreps setup      the setup wizard again: areas, topics, how often, how hard
npx atomicreps login      sign in from a browser with a typed code
npx atomicreps connect    register the server with your editor
npx atomicreps connect --pin   pin the launch args to this installed version, for a command you can commit
npx atomicreps mcp        the stdio MCP server process (what the editor launches)
npx atomicreps doctor     token, server ping, version, when the next rep may come, allowlist
npx atomicreps why        the last ten times a question could have appeared, and why it did or did not
npx atomicreps logout     forget the token on this machine
npx atomicreps logout --purge   also forget your reps and status, and offer to undo connect

Add --alpha to any command to use staging. It keeps its own token and cache, so you can be signed in to both staging and production at once.

Plain connect writes npx -y atomicreps mcp, which always resolves to whatever is latest at launch. connect --pin writes npx -y atomicreps@<version> mcp instead, the version this CLI is running, so an organisation can review one exact command and commit it rather than trusting npx to fetch the same thing twice.

Claude Code

npx atomicreps connect offers four Claude Code rows:

• Claude Code registers the server at user scope.
• …and a question at the end of each turn adds a Stop hook and a UserPromptSubmit hook to ~/.claude/settings.json. When Claude finishes a turn that changed files, the question prints in your terminal without Claude taking an extra turn, and the letter you type next is graded. Hooks you already have stay.
• …and allow its four tools adds rep, answer, me and settings to permissions.allow, so the first question does not stop on a permission prompt.
• …and a Claude Code status line shows the open question under the prompt. A status line you already have keeps printing first.

Without connect, the server alone:

claude mcp add --scope user atomicreps -- npx -y atomicreps mcp

GitHub Copilot

npx atomicreps connect offers VS Code and Copilot CLI as rows. Without it:

code --add-mcp '{"name":"atomicreps","type":"stdio","command":"npx","args":["-y","atomicreps","mcp"]}'
copilot mcp add atomicreps -- npx -y atomicreps mcp

Copilot in JetBrains, Visual Studio, Xcode and Eclipse reads the same entry from its own mcp.json (Settings, then MCP). A team can commit it as .vscode/mcp.json so every clone has it:

{
  "servers": {
    "atomicreps": { "type": "stdio", "command": "npx", "args": ["-y", "atomicreps", "mcp"] }
  }
}

Copilot reads the server's instructions and calls rep when a task ends, the same as Cursor and Codex. Approve the four tools once and they stay approved.

Cursor, Windsurf, Codex and other clients

connect adds one entry to ~/.cursor/mcp.json or ~/.codeium/windsurf/mcp_config.json, and runs codex mcp add for Codex. Anything already in those files stays. By hand:

codex mcp add atomicreps -- npx -y atomicreps mcp

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

Any other MCP client takes the same command: npx -y atomicreps mcp. The "I'll set it up myself" row in connect prints it and writes nothing.

How a rep is served

When your coding agent finishes a task, it calls rep once with a few words for what changed. The server picks one question, or nothing, and your agent shows it to you unchanged. Reply with a letter, or ignore it.

Under the verdict, one line names up to three other things the session touched. A number asks about one of them. "Never Swift" mutes a topic permanently. "Not this topic" mutes it for thirty days. "Unmute Swift" removes the mute. "How am I doing" prints your summary.

The server enforces the pace, not the agent. It sets a minimum gap between automatic reps, a daily cap, and mute and off. It refuses an answer typed within seconds of the question. It never includes the answer key in a rep. Every failure returns no rep rather than an error; doctor says why.

Tools

tool	what it does
rep({ touched?, ask?, topic?, kind?, exclude? })	one question, one insight, or nothing; ask takes a handle from an offer
answer({ pick, id? })	the verdict, the explanation, and the offer line
me({ show? })	summary, skills, reps, streak or mutes
settings({ intensity?, topics?, levels?, prefer?, mute?, muteMinutes?, days?, unmute?, dialog? })	the rate, the pinned topics, the difficulty range, the preferred areas, mutes, the dialog

Handles are topic or topic.subskill (react.hooks_core). Fifty sub-skill slugs repeat across topics, so the topic is always part of the name.

settings({ dialog: true }) asks for the letter in your editor's native dialog instead of the chat. It blocks the turn, so it is a setting, never the default.

Privacy

What leaves your machine on a rep call: catalog handles with small weights. Also package, file extension and folder names the public catalog already knows. Your agent's touched words are matched to handles on your machine; the words themselves are never sent. An unknown package name is dropped. A private name (@acme/react-internal) is sent as its public alias (react). We never send your code, a file's contents, or a prompt. npx atomicreps setup prints the exact payload for your repo, before you sign in.

npx atomicreps mcp forwards each JSON-RPC message from your editor to the Atomic Reps server with your token. The answer comes back the same way. ATOMICREPS_API is honoured only for an Atomic Reps or loopback address, since your token goes wherever it points. Anything else is ignored and reported by doctor; set ATOMICREPS_UNSAFE_ORIGIN=1 to override.

Login

npx atomicreps prints a code. Type it at atomicreps.com/connect while signed in; a forwarded link approves nothing. The token is saved in ~/.config/atomicreps/config.json (mode 0600), prefixed arep_ so secret scanners find it. You can revoke it from your account page. logout forgets it; logout --purge also deletes the reps and status this machine cached, then lists every editor entry connect wrote (both channels) and removes the ones you leave ticked.

The trust boundary, and how to report a vulnerability: SECURITY.md.

Check it yourself

dist/cli.js is built from these files and nothing else. The package has no runtime dependencies.

npm install
npm run build    # writes dist/cli.js from src
npm test         # the behaviour described above, as tests

To compare with what npm serves: npm pack atomicreps, then read its dist/cli.js beside the one you just built. Everything the client sends is in src/api.ts. Everything it reads from your repository is in src/infer.ts.

MIT.