If you work with AI agents

In short. Your agent already writes a log of everything it did. Habitat reads that log by rules written down in advance, shows you what it finds next to what it can't see, and gives you a signed receipt you can show someone else. Nothing has to be installed in the agent, and the log doesn't leave your computer.

The questions you might be asking

• Did my agent do something I told it not to?
• Did it touch files outside the folder I gave it?
• Can I show a colleague, a client or an auditor what it did, without handing over the whole log?
• Across a month of sessions, how often does this happen?

What runs for you today

One session, in your browser: Record. It is behind the Desk password until it is fully vetted. Drop in a Claude Code session log and press Check. Three checks run on it: two rules about what the agent did, and the file's fingerprint.

rule	the question	what it can't see
Repeat	Did the agent re-issue an action you refused?	whether you later told it to go ahead
Scope	Did it write outside the folder it was given?	anything done through a shell command

You get the session laid out as plain events, a dot on anything found, and a receipt signed by your browser's key. Try it in two minutes → · Record, step by step

All your sessions, on your machine: the census tools. The same checks over every log you keep, split into your own sessions and the subagents they ran, by month. You sign the result as a report, and you can choose to record its fingerprint in a public log so anyone can see when it was made. A demo shows what happens when someone edits a count in a report: the check refuses it. The code is not public; ask us if you want to run it. The census tools →

If you run agents for a business

If you run agents for a business, you probably already have to say what each agent may reach and which actions need approval. Habitat covers the part after the fact: what actually happened, read from the record, in a form someone else can check.

• Today: Claude Code logs, Repeat and Scope, single sessions in the browser, many sessions with the census tools.
• Not built: rules for other agents and log formats. Each one needs its own written rule, tested before it is used. The same goes for a rule that checks a team's own approval list against the log. If you need either, talk to us.

What it won't do

It won't say whether the agent was right. It won't judge a write outside the folder; it lists it. It doesn't watch an agent live or stop it. And it reads only what the log recorded.

Go deeper: What is kept, and where · The method, line by line · The agent-trace receipt