Docs

Before you start

• An account. Signup is open on this deployment — create one, and you are the owner of a new org.
• A computer on one of the platforms an installer is published for: Windows, macOS and Linux. The table below says which file is which. The native iOS and Android apps are in active development and not yet published for general download. Until then, one thing does work from a phone: signing in from any browser, including a phone browser, reaches machines, the agent, files and approvals.
• Node >=22.13.0, for the AI-client step only. The MCP server the app points your client at needs it; the desktop app, machine enrolment and the web app do not. The app checks for it and reports what it found rather than writing a client config that cannot run.

Setup, step by step

The desktop app has a Setup panel that performs and re-checks these one row at a time, so you do not have to hold the order in your head. Every row re-probes on each check rather than showing a cached answer, and most carry a re-run button for later. What follows is the same sequence, so you can read it before installing anything.

Step 1's files are the current release, 0.1.31. They are served without signing in, and SHA256SUMS.txt beside them lets you check the bytes you received — with the honest limit that the digest comes from the same origin as the download, so it proves the bytes arrived intact, not that this site is honest.

Which installer

The download page picks one for your browser and lists all of them; every link above is the same immutable, versioned URL it serves.

Which AI clients

The state beside each client is the manifest's own word for it, not a badge somebody typed. Connect your AI client has the exact command for each one and a handshake you can run yourself to prove the server answers before you blame a client.

Step 6 above does the wiring for 3 of these from inside the app, so the notes' references to scripts are the alternative route for somebody working from a source checkout, not a prerequisite for you.

What this documents

The versions these instructions were written against. Each is read from that component's own manifest, so a component that moves without this page moving fails the release rather than shipping a stale number.

Core ideas

• Org. Everything — machines, members, plan, agent traffic — is scoped to one org. You can belong to more than one and switch between them.
• Machine. A computer you enrolled. It dials OUT to Hawkeye over its own tunnel; nothing on the internet dials in to it. The agent does keep one listener of its own, on 8891, for callers already on your LAN or tailnet — a faster local path, same key as everything else, and XFER_PORT_DIRECT=0 turns it off.
• Engines. The coding agents the app can start on an enrolled machine: Claude Code, Codex or Grok. Each runs with the login already in that app; Hawkeye never reads the credential.
• Agent channel. How AI sessions on your machines talk to each other and to sessions in another org, when both sides approve the connection. Every cross-org call is a request the other side grants, not an obstacle to route around.
• Audit. Privileged actions are written down — who did what, on which machine, on whose authority. What that gives you between those boundaries is a record of what was done, not a gate in front of it. Read it as evidence after the fact, not as a control that stops something first. One caller stands outside even the record's attribution: the shared fleet key is a break-glass credential, accepted without an identity chain, so its actions are attributable to the key rather than to a person, and the policy that is being observed is not evaluated for it at all. Its uses are counted and the recent ones kept, which is how a machine can show you how often break-glass was reached for. Narrowing it is open work, not a shipped property — treat possession of that key as possession of the fleet.
• Authority (observe_only). Enrollment and client pairing require approval. After pairing, allowed actions execute without per-action approval: each machine's policy is evaluated and written to the audit trail in observe mode (XFER_POLICY_MODE defaults to observe, deliberately) rather than applied. A connected AI client can still raise an explicit human request when it creates one, and cross-org access is still a request the other side grants.

Taking it back

Three things can be revoked from the panel once you are signed in, and one cannot. Saying which is which matters more here than anywhere else on this page: a paired client is root-equivalent on every machine your org has enrolled.

• A paired AI client. /devices lists every app credential paired to the org and revokes any of them; a revoke takes effect on that client's very next call.
• An install link you sent someone. /installs revokes it, and a revoked link enrols nothing afterwards.
• A person. Removing a member deletes the membership and drops the org from their existing sessions in the same write, so their access to this org ends immediately rather than at a session's next expiry.
• A machine cannot be de-enrolled from the panel, and there is no uninstaller. Stated plainly because the alternative is a reader following an invented procedure: no route in this control plane removes an enrolled machine, and this repository ships no uninstall script for any platform. Stop the machine's services and remove its install directory yourself, revoke the paired clients above so nothing can reach it, and write in for the org-side record. That is the true answer today, not the one anybody wants.

Plans and billing

See the pricing page for what each plan includes. Once you are on a paid plan, invoices, card updates and cancellation are self-serve from the Plan page's "Manage billing" button.

Your data

See Privacy for what is collected and how to export or delete it (both are self-serve, from /account once signed in).

If a step does not work

• Re-run the single Setup row that is not green. Each one re-probes on every check and reports what it actually found, including "this platform does not measure that", rather than caching a stale answer.
• Ask the connected client for hawk_health — it reports the daemons, the job queue, the browser pool and desktop availability in one answer, so it tells you which link in the chain is down.
• Check status for the control plane itself. It is a live check run on the request that renders it, and it says out loud what it cannot observe.
• Still stuck: support. Send the step name and what the Setup row said — never a token, a key file or a raw privileged log.

Not written yet