When does the question appear?

After your agent finishes a task. It prints one question about the code it just wrote, under the agent's own last line, and a letter answers it.

How much does one question cost?

Your agent sends a topic and a few words about what it just touched, then prints the question it gets back word for word. Atomic Reps fetches a question that already exists; nothing is written at request time. Counted on the session replayed on the page with the o200k_base tokenizer: 15 tokens in, 171 out, about $0.0004 at Grok 4.6 list price (Sep 2026).

Can I change how often it asks?

Yes. You set how many a day and how far apart they come, and none at all is one of the choices. Ask your agent for a different rate and it changes the setting for you.

Claude Code · GitHub Copilot · Cursor · Codex · any MCP client

Know the code your agent wrote.

Atomic Reps Agent asks you one question about the code your coding agent just wrote, in the same session, seconds after it finishes.

npx atomicreps

Set it up for your agentCursor, Copilot, Codex, Zed

Or answer one in the browser

filter panel close animation is janky on staging, hangs then snaps. fix

⏺ Read(src/components/FilterPanel.css)

⎿ Read 61 lines

⏺ Grep(pattern: "max-height", path: src/components)

⎿ 2 matches

⏺ Update(src/components/FilterPanel.css)

⎿ Updated 1 file, +11 -6

⏺ Bash(pnpm test src/components/__tests__/filter-panel.test.ts)

⎿ 4 passed

⏺ Fixed it in src/components/FilterPanel.css. Want me to check the other panels for the same rule?

Atomic Reps · CSSREP 01

A panel holds 4000px of content and closes with the rule below. What does the box you can see do during the 300ms?

.panel { max-height: 9999px; transition: max-height 300ms linear; }
.panel.is-closed { max-height: 0; }

02About the change in front of you

The question comes from what your agent just changed.

Agents write code faster than you can build a mental model of it. Atomic Reps maps what your agent touched to one skill, then serves a question written in advance from that pool while you still remember the change.

the edit, and the rep it produced

⏺ Update(Dockerfile)

⎿ Updated Dockerfile with 6 additions and 5 removals

-FROM oven/bun:1.2-debian

+FROM oven/bun:1.4.2-slim

WORKDIR /app

+ENV PATH="/app/node_modules/.bin:$PATH"

-COPY . .

+COPY package.json bun.lock ./

-RUN bun install

+RUN bun install --frozen-lockfile --production

+COPY . .

-RUN apt-get update && apt-get install -y curl

EXPOSE 3000

-CMD ["bun", "src/index.ts"]

+CMD ["bun", "run", "src/index.ts"]

matched Docker · Dockerfile Best Practices

Atomic Reps · Docker · Dockerfile Best PracticesREP 01

You moved COPY package.json above COPY . . Which change to the repo now still makes npm install run again?

03Answer one

Pick a task your agent just finished.

Each tab shows the words your agent sends when it finishes and the question it gets back, printed the way your terminal prints it. Answer with a letter.

⏺ Moved the orders list from useEffect to useQuery in src/orders/OrdersList.tsx.

⏺ atomicreps - rep (MCP)(touched: ["@tanstack/react-query", "useQuery"])

Atomic Reps · State Management · TanStack Query / SWR · Level 4

What tradeoff does increasing staleTime introduce?

From memory. Reply with a letter.

04What gets asked

Written in advance. Not generated when you ask.

The question is not written in the moment your agent finishes. It was written before it was ever served, and fixed rules match it to the change. No model picks it and no model grades you. Report one that is wrong and it gets fixed. Answering from memory before you see the explanation helps you remember it later, and a question you miss is asked again until you get it right.

Atomic Reps does not use language models to process your answers. Inside a coding agent, that agent's own model reads your answer before it reaches us.

Read the research behind this

Never asked01/03: How many values does justify-content accept?
Asked: A panel holds 4000px of content and closes with max-height going from 9999px to 0 over 300ms. What does the box you can see do?
Never asked02/03: What does Promise.all return?
Asked: Five requests run in one Promise.all. The second rejects after 10ms and the rest resolve after 500ms. When does the await settle, and what happened to the other four?
Never asked03/03: What is the useEffect dependency array for?
Asked: An effect subscribes on mount, and its dependency array holds an object built during render. How many live subscriptions exist after three renders?

05How to set it up

Change any setting by asking your agent.

It happens in the session you already have open.

what you can ask for01settingsclaude

light

⏺ settings({ intensity: "light" })

⎿ Intensity light.

me now prints

Intensity light · 60 min apart

Three a day at most, and never two inside an hour.

Every reply here is the one the server sends.

06What it costs your agent

What a question costs.

One question for your agent, two ways to get it.

Generate it from scratch: research it, write it, source it, and check it.

Or fetch one of 140,000+ Atomic Reps, already written, kept current, and ready to use.

ATOMIC REPS · 15 tokens sent, 171 read. Half a second.
GENERATED BY AGENT · 996,244 tokens read and written, 839,194 of them its own transcript re-read from cache. 2 minutes to a draft; human review still to come.
Both measured, Sep 2026, for one question: the generation as one run of a Claude Opus 5 agent at medium effort, the fetch on our server.

Model cost per question

What one question costs, by model
Model	Generated by agent	Atomic Reps
Grok 4.6	$0.52	$0.0004
Gemini 3.1 Pro	$0.57	$0.0005
GPT-5.3 Codex	$0.53	$0.0005
Claude Opus 5	$1.38	$0.0012
GPT-6 Astra	$2.75	$0.0025
At Sep 2026 list prices. The generated column is the measured run priced at each model's own rates, cached re-reads at a tenth of the input rate, and leaves out the human review that follows. Atomic Reps fetches the same size whichever model you run.

Before the agent drafts one, Atomic Reps serves 240.

what 240 questions cost, against the clock01a day of questions

The server serving each question ATOMIC REPS
An agent drafting each question, one measured run GENERATED BY AGENT

07 · Retrieval practice, measured

AI can give you the answer.Understanding it is still your work.

None of these studies tested Atomic Reps. Read every study and its caveats

08Connect

Connect the coding agent you already use.

In the terminal

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

User scope, so it is there in every repo you open.

Then what

Run it once. It writes one server entry and nothing else.
Open Claude Code in the repo you are working in.
Finish a task. One question about what just changed arrives in that session.

In the terminal

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

Or skip the terminal:Add to VS Code →

The first time Copilot calls a tool it asks; pick Always and it never asks again.

Or the file · .vscode/mcp.json

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

Commit it and every clone of the repo has it.

In the terminal

copilot mcp add atomicreps -- npx -y atomicreps mcp

Then what

Run it once. It writes one server entry and nothing else.
Open Copilot in the repo you are working in.
Finish a task. One question about what just changed arrives in that session.

Copilot · Settings · Model Context Protocol · mcp.json

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

Copilot in JetBrains, Visual Studio, Xcode and Eclipse reads the same entry VS Code does.

Then what

Paste it into the file. One server entry and nothing else.
Open your editor in the repo you are working in.
Finish a task. One question about what just changed arrives in that session.

~/.cursor/mcp.json

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

Or skip the file:Add to Cursor →

Then what

Paste it into the file. One server entry and nothing else.
Open Cursor in the repo you are working in.
Finish a task. One question about what just changed arrives in that session.

In the terminal

codex mcp add atomicreps -- npx -y atomicreps mcp

Then what

Run it once. It writes one server entry and nothing else.
Open Codex in the repo you are working in.
Finish a task. One question about what just changed arrives in that session.

Any other MCP client

POST https://atomicreps.com/mcp
Authorization: Bearer <your token>

The token is minted on your account page and shown once. Revoke it there and the connection closes the same second.

Then what

Paste it into the file. One server entry and nothing else.
Open your client in the repo you are working in.
Finish a task. One question about what just changed arrives in that session.

One server. npx atomicreps connect configures the client you choose.

Work in VS Code? The extension asks about the code you just committed, without leaving the editor.

09Try it

Try it in five minutes on a repo that is broken on purpose.

broken-by-design is a small support-desk app with twelve planted bugs. Start it with Docker, connect your agent, and ask it to fix whatever you notice first. When the fix is done, you answer a question about the code it changed.

Open the repo on GitHub

AtomicReps

Run it in the repo you have open

The next time your agent finishes a task, answer one question about what changed.

npx atomicreps connect

Create a free accountno card required Approve your terminalalready connected