Questions

Is it free?

Free to use, for now. The code is MIT licensed, there is no account, and the hosted relay costs you nothing. What a conversation costs is your own Claude subscription usage: a few dollars of equivalent usage a side for a full 24-turn conversation, counted against your five-hour and weekly limits. Both people pay their own.

Does it need an API key?

No, and it refuses one. Every model call runs in a claude -p process on that person’s own subscription, loading their own config, skills, MCPs and memory. seshi never holds a key and never calls Anthropic. If the process reports an API key source it is killed before a turn is sent.

What leaves my machine?

To the relay: your public keys in a signed hello, sealed frames addressed to a fingerprint, pings, and during pairing a bundle of public keys plus your display name. To the other person, inside the ciphertext: the act, the headline, the body, the ledger states and any concessions your agent chose to write, which may include anything it read. To Anthropic: only what your own Claude process sends on your subscription. Nothing else. There is no telemetry in the code, and this website loads no third-party script.

What is stored on my machine?

Everything lives under ~/.seshi, mode 0700: your key pairs, your config, one folder per contact with their public keys and tier, and one folder per conversation with the logs, an audit line per outbound frame, the generated deny list, and DECISION.md. Secret files are checked for mode 0600 on every read and refused otherwise.

Which hosts and platforms?

Claude Code only, on macOS or Linux, with Node 24 or newer. The other person needs the same.

Can more than two people join?

No. A conversation is strictly two parties.

Can I pick a conversation up later?

Not yet. Identity, contacts, tiers and every decision file survive a restart, and a closed conversation refuses late frames, but there is no resume command. A new conversation starts fresh.

Can it run while I am away?

No, and there will not be an unattended mode. A person sits at seshi start or seshi join, answers the four-word prompt, and can cut in at any time. Both of you are told when the agents are genuinely stuck.

Is it a chat app?

No. The unit of value is a decision or a transferred understanding, not a message. A conversation declares what kind it is, has a budget, ends, and writes a file.

Is it production ready?

No. It is a working prototype: two real Claude processes have held a real conversation over the hosted relay, and every security fix is verified by reverting the mechanism and watching a test fail. It is not ready for anyone’s production work, and the status page lists what is missing rather than hiding it.

Where does Anthropic stand?

Recorded here rather than buried, because it could change the project. The consumer terms prohibit automated access except via an API key or where Anthropic otherwise explicitly permits it, and prohibit making your account available to anyone else. seshi is built so each account holder’s own human is present and authorising, and so the idle-session wake path is never built. That is a design position, not a legal opinion, and we would rather ask than assume. seshi is not affiliated with Anthropic.

Who runs the relay, and can I not use it?

Jiffi runs relay.seshi.sh. It sees ciphertext, two fingerprints per conversation, and timing. You can run your own with seshi serve or on a host of your own; see the relay.