JQ Groundcrew
The setup check for Claude Code.
A coding agent works inside a setup you made once and rarely look at again: an instructions file, a list of approved commands, a few settings. Groundcrew reads that setup and tells you what to fix. Each fix comes with a prompt you can paste into Claude Code.
Run it
In the folder of a project, with Node 18 or later:
npx jq-groundcrewThen open http://127.0.0.1:4317. Refresh the page to check again after a change. Add --json to print the report, or --html report.html to save the page.
In CI
From version 0.1.1, --fail-on fail makes the command exit with code 1 when any check fails, so a CI step needs one line. --fail-on warn also counts a warning. The report still prints in full, and each counted check is named on its own line.
npx jq-groundcrew@0.1.1 --json --fail-on failPin the version, so a new check cannot fail your build overnight. Running it only when the setup changes (CLAUDE.md, .claude/, .env files) is enough.
Not every check means something in CI. A fresh clone has no .env file, because those are ignored by git, so the secrets check passes there whether or not a rule protects them. Run that one on a developer’s machine. In CI the useful checks are the ones about files in the repository: CLAUDE.md growing past its target, a hook being removed, and a rule that approves every command. DeckSketch, one of our sister products, found this when it added the step, and suggested we say so here.
What it checks
Setup
What loads into every session, and what runs without being asked?
- The length of each CLAUDE.md, against the documented target of under 200 lines
- Whether the memory index is longer than the part that loads
- Whether any hook is configured
- Whether any command is approved in advance
Guardrails
What may the agent reach, and what stops it?
- Whether a deny rule keeps the agent out of secret files such as .env
- Whether a rule approves every shell command
- The permission mode
- Whether the sandbox is on
Verification
Who checks the work when the agent says it is done?
- Whether the agent can run a test, lint or type-check command without asking
- Whether a hook runs when the agent finishes
- Whether each subagent is limited to the tools its job needs
Each check names the page of Anthropic’s Claude Code documentation it rests on, so you can verify the advice. Advice that is ours, and not the documentation’s, says “our suggestion”.
What a result looks like
This example is a made-up project. The numbers are invented to show the page.
FixSecrets are off limits
No deny rule covers secret files, and this project has .env and api/.env. The agent can read them.
FixCLAUDE.md length
410 lines. The docs say to target under 200 lines per file. Every session loads all of it.
LookSomething checks the work before the agent stops
No hook runs when the agent finishes. The agent decides on its own that it is done.
GoodNo blanket approval of shell commands
No allow rule approves every shell command.
Each row that needs a fix comes with a prompt. For the first row above, it would read:
Add deny rules to the permissions.deny array in .claude/settings.json so that Claude Code cannot read secret files: Read(./.env), Read(./api/.env). Keep every rule that is already in the file. Show me the diff before you save it, and tell me how to confirm afterwards that a read of one of these files is refused.A prompt that edits a file always asks to see the plan or the diff first. Groundcrew itself changes nothing. You paste the prompt, you read what Claude Code proposes, and you decide.
What it reads, and what it never reads
It reads
- The Claude Code settings files: yours, the project’s, and any managed ones
- CLAUDE.md files, to count their lines
- The names of your skills, subagents, commands and MCP servers
- The names of files that usually hold secrets, to tell you whether a rule protects them
It never reads
- The contents of a secrets file
- A session transcript
- Your source code
- The network: it makes no outbound call of any kind
The page it shows you lists every file it read. It is one small program with no dependency, so you can read all of it before you run it.
What it found on our own repository
This part is measured, not an example. We ran Groundcrew on the repository behind this site on November 14, 2026.
- Our instructions file was 452 lines. We cut it to 172 by moving six sections into files that load only when needed.
- Four secret files had no rule keeping the agent out of them.
- No hook was configured, and the sandbox was off.
You can run the same command on your project and compare. Two things we learned while fixing ours are in a pre-flight checklist for a new Claude Code project and the Claude Code sandbox is off by default.
Limits
- It covers Claude Code only.
- It checks setup. It does not look at how a session went.
- It is new. The first version was published on November 14, 2026.
- The checks follow the documentation as it stood that day. If a check and the documentation disagree, the documentation is right.
- One check looks for a test or lint command by its name. A project with an unusual command name will be told it has none.
The board, for more than one project
We also keep a private board that shows a one-line summary for each project on each of our machines. Only the summary leaves a machine: the score, and the title and status of each check. What was found, the fixes and the prompts stay where they were made.
The board is ours alone today. Accounts for other developers are not built yet. When they are, the board will be part of Hive Platform Plus, and the command above will stay free.
JQ Groundcrew is made by Point Seven Studio. It is unofficial: not affiliated with or endorsed by Anthropic. “Claude” and “Claude Code” are Anthropic’s names for their products. The package is jq-groundcrew on npm.