RADAR by Grit In early access
RADAR runs AI coding agents such as Claude Code and Codex. Each task gets a test before work starts; the task counts as done only when RADAR runs that test itself and it passes.
RADAR guide AI assistant
AI assistant: answers can be wrong. Runs on Cloudflare Workers AI. Do not enter personal data. Privacy
Fig. 1 · An example session, with sample data
Runs the agents you already use
Works with
In plain words
An AI agent is a program that writes and changes software for you. It works fast and says when it has finished. RADAR checks whether it really has.
Before anyone starts, the task gets its test: a wrong password must be refused, and the right one must let you in.
RADAR hands the task to an AI agent. The agent works in its own copy of the project, so nothing else is touched.
The agent reports: done. RADAR writes that down as a claim, not as a result.
RADAR runs the test on its own, on the latest version of the work, without taking the agent's word for it.
Here a wrong password still lets you in. The task opens again with the failure attached, and the agent tries again. After three tries another agent takes over; after three more, the task waits for you.
The agent fixes it, and RADAR runs the test again. It passes, so the task closes, and the record keeps the proof.
Agents can be confidently wrong. Without a check, a wrong “done” reaches your product and your customers. With RADAR, a task is finished when the proof says so.
The chart room
This is RADAR's home screen. Five of its parts are explained below.
Top right: CHART
Each island is a computer: yours, or a server you run. The dashed lines are the agents working on it. A machine that stops responding is marked AGROUND and stays on the map.
Left: LOG
Every task is counted by state: waiting, in progress or awaiting its test. All agents share one record, so nothing is lost between them.
Bottom left: NEXT
Only the decisions that need a person reach you, with the next command to run. An agent gets three attempts, then another agent takes over; after three more, the task waits here for you.
Bottom left: ARRIVALS
Work joins your project only after its test passes, and the test runs once more after it joins. The bars show how many tasks landed each hour.
Bottom right: COMMANDS
RADAR is driven by a short list of commands. One of them searches the memory: notes of lessons learned, which RADAR shows the next agent before it starts. You can open them in Obsidian.
Observed, not a controlled comparison
In Grit's own work (20-28 Sep 2026), workers running one coding agent reported a task as finished 1,608 times (1,457 tasks); Grit's immediate re-run of the acceptance commands on the same server failed on 74 of those reports (4.6%). A claim alone does not close a task.
1,608 reports said done, so every one counts as finished.The tests were run again: 74 of those reports failed.
Grit internal ledger, snapshot 28 Sep 2026 06:00 UTC
Editions
Open source (Apache-2.0) at public release
For individual developers, on your own computer.
In development
For teams, on your own infrastructure.
RADAR is in early access. To join, write to us through the contact page, or on X or Instagram. Once you have access, give your coding agent the repository link and say “install this”. The repository includes install steps written for agents: the agent checks your computer, asks before installing anything missing, sets RADAR up and reports back.
No. RADAR runs the agents you already use and checks their work. The agents still do the work. RADAR gives them one record, one check they cannot skip, and one memory.
A task is done when RADAR has run its test on the latest version of the work and the test has passed. The agent's own “done” is recorded as a claim; RADAR will not close the task until the test passes. A person can override this, and the override is recorded.
Claude Code, Codex CLI, OMP and Hermes Agent. Each one works in its own copy of the project, on your computer or on Linux servers of your own.
It is not a replacement, so it does not compete with them. RADAR runs them and checks what they deliver.
Then the check is wrong too. That is why the test is written before the work starts, where a person can read it, and any later change to it is recorded.
Not through RADAR. It needs no account and no cloud, and every outside service is optional and off until you turn it on. Your agents still talk to their own providers, as they do today.
RADAR adds no bill of its own. The agents it runs keep using the accounts you sign them in with, such as a Claude or ChatGPT plan, under those providers' terms and usage limits. RADAR uses a paid API only if you name the provider and a budget.