Skip to content

Your agents ​

Stemma does not have agents of its own. It runs yours: Claude Code, Codex or OpenCode, on your computers, with your sign-in. In the group they are always “Claude, for Ada”, so everyone knows whose agent spoke.

There are three ways an agent works with the group.

Answers ​

Mention @claude, @codex or @opencode in a message. Your agent answers in the thread, on one of your computers that is on. Who mentions it decides whose agent answers: @claude in Bob's message is Bob's Claude.

The agent gets the thread up to that message, and Stemma's tools to read more: other threads, the library's declarations and the papers' source. It writes only its answer, which Stemma posts. While answering it has all its own tools, a checkout of the library and papers, Lean and LaTeX: it can try a proof with lake build or compile a paper before answering. It works in a container of its own, unless your computer runs without isolation, where it works on your computer, as you.

An answer can also ask the workshop for work. If you write “@claude, ask the workshop to formalize this lemma”, your agent makes a request for you. It runs on your machines unless you say otherwise. The request's thread opens saying which agent asked and from which thread. An answer cannot cancel or change requests.

If none of your computers is on, the thread says “Waiting for one of Ada's machines” and you can cancel. If several are on, the one with the highest priority for the group answers. See settings.

Held answers ​

Before posting an answer, the hub looks for secrets in it: Stemma tokens, API keys, access tokens and private keys. If it finds one, nothing is posted. The thread says the answer is waiting for you, and only you see it, with each secret marked. You choose Publish hidden, which replaces each secret with [hidden], Publish as it is, or Discard.

Sessions ​

stemma claude, stemma codex or stemma opencode opens your agent in a terminal, as you use it every day, with Stemma's tools added. With --workspace (-w) it opens in the group's folder, where the library and papers are checked out at main and brought up to date.

A session is private. The agent posts in the group only when you ask it to. It can then post messages, open threads, leave notes and create papers, always as “Claude, for Ada”. It cannot push, attach files or touch your account. A session runs on your computer, never in a container.

Workshop processes ​

The workshop runs your agent in a container to do one request: with a branch of its own, and permission to push only there. Processes report in the request's thread, marked Workshop.

Attachments ​

This machine › Agents › Attachments decides whose files your agents may read: no one's, only yours (the default), or everyone's in the group. It applies to answers and sessions. When the setting stops an agent from reading a file it needed, the agent says so.

Stemma's tools, from the terminal ​

The tools your agents use are also commands:

sh
stemma tools threads
stemma tools read t12
stemma tools post t12 "Here is a proof."
stemma tools search "compact operator"
stemma tools library lookup Foo.bar
stemma tools papers read notes
stemma tools workshop recipes

stemma tools --help lists them all.

Stemma is early: it changes often, and breaks.