Disclaimer

Flanner scores how likely a plan is to have drifted from the code it describes. It does this from evidence it can see: commits, file changes, timestamps.

It cannot read intent. A plan marked fresh may be wrong, and a plan marked stale may be perfectly accurate. Treat the score as a reason to look, never as a reason not to.

Marking a plan accepted records that a person approved it. It is an audit trail, not an assurance that what was approved is correct, safe, legal, or a good idea.

Flanner is not a backup service. Plans sync between devices you control, and we hold no copy. If every one of your machines is lost, so are the plans, unless you have a backup of your own.

Use git, or your normal backups. The plan directory is an ordinary directory of Markdown files.

Every version is signed by the device that wrote it, and any device can verify that history. That proves who wrote something and that it has not been altered since. It says nothing about whether the content is right.

Flanner stores and versions planning documents your AI tools produce. It does not check them. Anything an agent writes is the agent's output and your responsibility to review, exactly as it would be without us.

Flanner Curb reads your agents' settings. It reports what an agent could reach if started the way the report states. It does not inspect running sessions, so a session started another way may reach more. A rating, a channel marked controlled, or an applied fix describes settings. It is not a guarantee of behaviour.

flanner curb test asks the agent to read a decoy file four ways. A pass proves only that this agent version refused those methods, on that target, in that launch, with those settings, at that time. It proves nothing about another file, method, launch or version. "Inconclusive" and "not tested" prove nothing. A pass in a scratch copy of a project is not a pass for the project.

Curb is tested against the agent versions in its support matrix. For any other version every result is an assumption.

Curb is not a sandbox, a network filter or a secrets vault. It does not protect against software running as you, including an agent that writes and runs its own code.

A test uses your own model provider account and costs tokens. The test prompt and the decoy's fake contents go to that provider.

Nothing in the product or on this site is legal, financial, security or other professional advice.

The client is provided under the MIT licence, which disclaims warranty in its own terms. Read it in the repository.

Liability for the hosted service is addressed in the Terms of Service. Questions to support@flanner.io.