Tokens and trust
Stemma lets other people's requests run on your computer, and agents act in a group for you. What keeps that safe is that every piece of work carries a token made for it alone, and the hub checks every request against that token, not against what the agent says it is.
The tokens
| Token | Who holds it | Lives | May |
|---|---|---|---|
| Personal | Your computer, or a script | Until revoked | What you may, as you |
| Agent's | An agent in a session | The session, a day at most | Read one group; post, open threads, leave notes, create papers, ask and cancel workshop requests, as “Claude, for Ada” |
| Job's | An agent answering | The answer | Read one group; ask the workshop for you. The hub posts the answer |
| Process | A workshop process: its script and its agents | The lease | Read the group; post and attach in its request's thread; what its recipe's permissions allow; fetch the group's repositories; push only its own branch |
| Build | A build | The lease | Fetch one repository; report that build |
Your personal token never reaches an agent, and neither does any token made in Settings › Devices. Agent, job, process and build tokens cannot reach your account, other tokens, invitations, members, settings or administration.
The tools an agent sees are exactly what its token allows. An answer's tools are read-only plus ask_workshop. A session's add posting. A process's let it post in its own thread, and do what its recipe's permissions allow: posting elsewhere, opening threads, leaving notes or creating papers. A process never does more than its requester may.
Git
All git traffic goes through the hub, which checks the token and the group's roles before passing it on to Forgejo. For a process, the hub reads the push's command list before streaming it: a push to any branch other than its own, workshop/<request>-<process>, is refused before it reaches Forgejo; a process whose recipe opens no pull requests pushes nothing. Only admins and the reserved workshop account may write main, and workshop merges only what a person approved.
Approval on the machine
Nobody can make your computer run someone else's work but you. A workshop offer from another person waits on your computer until you approve it there, in the app or with stemma workshop approve. Approvals are kept in a file on your computer, not on the hub. The hub cannot approve, and neither can a hub admin. Before starting, your computer checks the offer again against the approval: the same recipe version, the same requester, and within its limits.
Containers
With isolation, an answer or process runs in a container that:
- runs as you, never as root, with no extra capabilities, and with limits on memory, CPUs and processes;
- sees only the group's folder (answers) or its own checkouts (processes), its job's folder with its token, read-only, and the agent's container sign-in;
- never sees your own sign-ins to Claude Code, Codex or OpenCode, your Stemma credentials, or any
STEMMA_*variable; - is removed when the work ends.
Builds run in containers too, with no model sign-in at all.
Without isolation (machine.isolation = "none"), answers, processes and builds run on the computer, as its person, with nothing of the above: everything that person can reach is within the agent's reach. It is the owner's choice, and stemma doctor advises against it.
Secrets in what agents write
Before posting an answer, the hub scans it for secrets: Stemma tokens, Anthropic, OpenAI and GitHub keys and tokens, JWTs and private keys. A match holds the answer for its person alone. Workshop progress, process summaries and build logs have secrets hidden before they are stored.
Sessions are not a wall
A session runs your agent on your computer, as you. Stemma removes its own sign-in from the session's environment, and tells the agent to use only Stemma's tools. Even so, an agent that insists can read your files. Containers are the boundary; sessions are a convenience.
The group's model key
The orchestrators' API key is encrypted in the database with AES-256-GCM, under a key file the hub keeps beside it. It is never sent to computers, processes or prompts, and it is hidden in provider errors. The hub must decrypt it to make calls, so the hub's operator could read it. Use a key with a spending limit.