
Atomicreps
io.github.atomicrepsv0.1.0Updated Oct 1, 2026
Retrieval practice in your coding agent: one short question about what you just built.
Overview
Adds retrieval practice to a coding agent: after a task, the assistant asks one short question about what you just built.
- What it does
- When a coding agent finishes a task it calls rep once with a few words for what changed; the server picks one question, or nothing, and the agent shows it unchanged (R18, R19). Tools are rep, answer, me and settings: rep serves a question, insight or nothing, answer returns the verdict and explanation, me reports summary, skills, reps, streak or mutes, and settings controls intensity, topics, levels, preferred areas, mutes and dialog (R33-R40). The server enforces pace itself with a minimum gap, a daily cap, mute and off, refuses answers typed within seconds of the question, and never includes the answer key (R27-R30).
- When to use it
- Worth adding when you want lightweight, spaced retrieval practice about the code you just wrote, inside the editor you already use, without leaving the chat. It suits developers who want occasional short questions on topics such as React or Swift and who want to control how often they appear. It is not a code review, documentation or search tool.
- Requirements
- Node 24+ and a free account (R4). Runs as a local stdio process launched with npx atomicreps mcp, or via the remote endpoint at atomicreps.com; no environment variables or headers are declared in the manifest. Sign-in uses a typed code at atomicreps.com/connect (R56, R57). Works in Claude Code, GitHub Copilot, Cursor, Codex or any MCP client (R3).
Installation
In SourceWeft
- Open Atomicreps in the dashboard and add it to a workspace.
- Enable the server for the chats that should use its tools.
Web executable via Streamable HTTP. Remote servers run from the web runtime once configured in a workspace.
Other MCP clients
Add this to your client's mcpServers config.
{
"mcpServers": {
"atomicreps": {
"type": "http",
"url": "https://atomicreps.com/mcp"
}
}
}README
atomicreps
One short question about the thing you just built. Works inside Claude Code, GitHub Copilot, Cursor, Codex, or any MCP client.
Requires Node 24+ and a free account. On an older Node, every command prints the version it needs instead of starting.
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.
GitHub Copilot
npx atomicreps connect offers VS Code and Copilot CLI as rows. Without it:
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:
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.
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
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.
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.
Source: README.md at commit df3660b
Tools
0Version history
1- v0.1.0LatestOct 1, 2026