SSH made agentic.
Not made reckless.

1 credit = 1 agent turn: one message to the copilot, including the commands and follow-ups the agent runs inside that turn. A long task can cost a little more — the gateway adds a credit for every couple of extra model calls a single turn needs. SSH, the terminal, and SFTP are never metered.

For agent chat, yes — signing in is what mints the credits the copilot runs on. Everything else works without one: hosts, terminals, SFTP, and keychain storage are all local and never metered.

Hosts, chats, and history live in SQLite under ~/.tryssh. SSH keys, passphrases, and passwords live in the macOS keychain — never in a database, never sent to us. Your terminal sessions go straight from your Mac to your server; we are not in that path.

Agent chat is the exception, and we would rather you heard it here than found it later: when you send the copilot a message, that turn's content — your prompt, the commands the agent proposes, and the output those commands return — is relayed through our gateway to Anthropic to generate the reply. So yes: if the agent reads your nginx.conf, that file's contents leave the machine. We do not train on it. Delete ~/.tryssh and the local side is gone.

Exactly what leaves your machine, line by line →

No. The risk gate is default-deny: a command auto-runs only if it matches a read-only allowlist and contains no shell metacharacters — no pipes, redirects, semicolons, backticks or command substitution, the characters that let one harmless-looking command smuggle in another. Anything else, including anything that mutates state, renders the exact command and waits for your approval.

Agent chat stops and says so — you get an out-of-credits message, not a silent failure. Terminals, SSH, and SFTP keep working; they are never metered. Credits do not roll over: a paid allowance resets at the start of each billing cycle, and the Free plan's 200 is a one-time starter grant, so with no subscription nothing refills it. There is no top-up, no annual billing, and no trial of a paid plan today. If you exhaust Pro's 2,400 mid-month your options are upgrade or wait for the reset.

Any time. Access continues to the end of the period you already paid for, then the subscription stops. We do not publish a refund policy — if a charge looks wrong, email hello@definable.ai and a human will read it.

Not in the shipped Mac app today. The core supports direct Anthropic auth — it is how we develop it — but the app has no place to enter a key, so agent chat goes through our gateway on credits. If you want a bring-your-own-key mode, tell us at hello@definable.ai; it is the request we act on fastest.

Your hosts, chats, and history are SQLite files on your disk and your keys are in your keychain, so nothing of yours is locked in a service. What you would lose is the hosted copilot, because agent chat runs through our gateway. We are a small team and would rather you sized that risk yourself than found out later.

macOS only today. The core is a portable TypeScript backend behind one WebSocket protocol, so the brain already ports and Windows and Linux clients are on the roadmap — but neither is shipped and we are not going to give you a date we cannot keep. Leave your email and we will write to you when there is a build.

Email me when it ships

Notify me

Windows Linux

One email when there's a build. Nothing else, no list.