2.0 · open source · MIT

A team brain for AI coding agents. It makes the agent check its work, and makes you better at directing it.

Plain files in git and Claude Code hooks. No account.

npm install -g codehere
codehere init

codehere

scroll

01 · What it does

From vibe coder to vibe engineer.

You keep using Claude Code as usual. Codehere talks to the agent, not to you, and stays silent unless it helps. It puts five engineering habits inside the agent:

You ask for a feature with no test

The agent first asks you one question: "How will you know it works?" It is told to turn your answer into a failing test before it writes code.

You type "still broken" or "masih error"

Your short message reaches the agent with the real error, the failing test and the files changed this session.

The same error comes back 3 times

The agent is told to stop patching and name a root cause before it edits again.

The agent says "done"

Codehere runs your tests itself. If they fail, the agent goes back to work. A skipped or deleted test is blocked.

You are still stuck

codehere help writes a redacted help card for a teammate. Their reply --solved becomes a team lesson that every teammate's agent gets the first time the same error appears.

npm install -g codehere
cd your-project
# finds your test command, adds the hooks for the agents you name
codehere init --agent claude,codex,copilot

codehere rule "Prices are in cents. Never use floats for money."
# a help card for a teammate
codehere help "Cart total ignores quantity"
# their answer becomes a team lesson
codehere reply <id> "Use price * qty" --solved
# one clean spec for a fresh session
codehere brief

02 · How it decides

Plain rules decide. No model judges a model.

Loops, failed tests and skipped tests are found with plain rules, so Codehere does not inherit the agent's mistakes. It only reacts to what changed in this session, stays silent on normal turns, never blocks the agent more than 3 times in a session, and if anything breaks it exits quietly and the agent carries on.

The team brain is plain markdown in .codehere/: rules written by people, lessons, help cards and restart briefs. It travels with your repo through git. Gemma 4 through Ollama is optional and only writes summaries for briefs and help cards, which you read before using.

03 · What it does not do yet

Its limits, written down.

Claude Code, Codex CLI and GitHub Copilot CLI only. On Copilot, and in codex exec, you do not see Codehere's one-line notes; the agent still gets every message.
Lessons match by error text, so a very generic error can bring up a lesson that does not fit. The agent is told it "may apply".
The weakened-test scan catches skipped tests, deleted test files and removed assertions. It does not yet catch an edited expected value.
Redaction catches common keys, tokens and emails. Read a help card before you share it.
Sharing is through git. There is no server, so teammates see updates after git pull.

04 · FAQ

Frequently asked questions

›What is codehere?

A team brain for AI coding agents: plain markdown files in your repo plus Claude Code hooks. It stops the agent going in circles, checks "done" by running your tests, and shares what your team learns. Open source, MIT.

›Which agents does it work with?

Claude Code, Codex CLI and GitHub Copilot CLI, through their hooks. The CLI and the team files work with any tool.

›What happened to codehere check?

It is still on npm as version 1.x: run npx codehere@1 check. The GitHub Action pins its own 1.x version, so it keeps working.

›Does my code leave my machine?

No. Everything is files in your repo. Help cards are redacted, and you choose whether to commit them or paste them into a chat.

›How am I charged?

You're not. codehere is free and open source. No account.

Source and README: github.com/muhammadegaa/codehere