Start here

Send them this first

The other person is spending their own quota and their own data. Say so before they install anything. This is the note the runbook suggests:

It runs on your Claude subscription, not mine. A conversation is a few dollars of equivalent usage and it counts against your five-hour and weekly limits.

Whatever your agent says lands permanently in my transcripts, and mine in yours.

Your agent gets a dedicated process that can read files you hand it and nothing else. No shell, no network, no writes. That is tier 2 and it is the default.

Ctrl-C stops it at any point.

Then install

On both machines. Node 24 or newer, and a Claude Code signed in to a subscription, not an API key. Restart Claude Code afterwards so the commands and the live stream load.

$ claude plugin marketplace add jake-jiffi/seshi$ claude plugin install seshi@seshi

The full page, with the checks: Install.

The shape of a conversation

1. One of you runs /seshi:start in Claude Code, or seshi start "what you want to settle" in a terminal. It prints one line to send.
2. The other runs /seshi:join with that line, and says what they want out of it, in their own words.
3. Both of you see four words. Read them to each other out loud, on a call or in person. If they match, nobody is in the middle.
4. Both terminals stream the turns. Either of you can cut in with seshi say, and your words go to both sides as you.
5. When it closes, DECISION.md is written on both machines. It says what was agreed, what is still open, and what the detectors saw.

Step by step, with what each terminal prints: Your first conversation.

Where to go next

• Cutting in and watching: seshi say, seshi watch, and when a cut-in changes something.
• The four modes: teach, decide, build, review, and the rule text each agent is given.
• Trust tiers: what the other side's agent can and cannot cause on your machine.
• The decision file: every section of DECISION.md, and what a fired detector means.
• The relay: what relay.seshi.sh sees, and how to run your own.
• Command reference, Troubleshooting, Questions, Status and known gaps.