Saeta is in development and has no public release yet. This page describes the current development build and our plans.
How Saeta decides
When an agent wants to run a command, edit a file or read one, Saeta checks its rules before anything reaches you. The first rule that matches decides. When no rule matches, Saeta asks you.
The order
Saeta reads the rules in this order:
- Your organisation's blocks. Only on a computer an organisation manages (see For administrators). Nothing below can let through what one of them blocks.
- Saeta's own protections. Hard blocks, such as
rm -rf /, a force push, a download piped into a shell, and any command that names your private keys or Saeta's token. And questions that always ask: Saeta's own files, stopping Saeta or running its CLI, an agent's hook settings, git's hooks and settings, shell start-up files. A rule of yours that says "always allow" can't get past them, and neither can an organisation's allow. - Your organisation's other rules: what it asks about, allows, or lets run in safe auto mode.
- Your rules. The answers you chose to remember, in the order the Trust List shows, then the exact commands in
dev_allowlistin~/.saeta/config.toml. - Your organisation's defaults. Suggestions from the organisation that a rule of yours can override.
- The rest of Saeta's starting rules (the seed). Reading a secret-looking file such as
.envasks; plain read-only commands such asls,cat,rgandgit statusare allowed.
Nothing matched: Saeta asks.
Without an organisation, steps 1, 3 and 5 are empty. Saeta's starting rules ship with the app and change only with a new version; their changes are in the release notes. You can switch one off in the Trust List or with saeta rules disable <id>, for example seed.deny.key-material if an agent really should handle your keys. saeta rules lists every rule with its id and where it comes from.
Two modes change what happens after a match, not the order: safe auto mode and agent decides.
What "ask" means
The agent waits. The request appears in the desktop inbox, and on your phone if you paired one. You allow or deny it, and can tell the agent why. You can also remember the answer as a rule; see Trust and activity.
From a terminal:
saeta status # pending requests and their ids
saeta approve <id> # allow once
saeta approve <id> --always # and remember it for this program in this repo
saeta deny <id> --reason "not now" # deny, and tell the agent why
--exact remembers only this exact command; --everywhere remembers it in every repo. When an agent runs saeta itself, Saeta asks you first: the CLI can answer requests and change rules.
When nobody answers
A request waits up to default_timeout_secs (300 seconds by default, at most 300), then fallback answers it: "deny" by default, or "allow".
# ~/.saeta/config.toml
default_timeout_secs = 300
fallback = "deny"
Some rules are high-risk. A request one of them sent to you is denied when nobody answers, whatever fallback says. Saeta's own protections that ask are high-risk, and so is everything safe auto mode still asks about. An organisation can mark its own rules high-risk too.
When Saeta isn't running
If the agent can't reach Saeta, daemon_down_behavior decides: "fail_open" (the default) lets the call through with a note, "fail_closed" blocks it. This is why stopping Saeta is one of the things that always asks.
Folders Saeta watches
By default Saeta asks about agents working anywhere. With folders in config.toml, it only asks about agents working in those folders. Elsewhere it steps aside, and the agent's own permission settings apply.
folders = ["/Users/me/work"]
What the rules can't see
Saeta reads the text of a command and the paths a file tool names. It does not see inside a script or a program: bash scripts/build.sh is judged as running that script, not by what the script does. A process running as you can also change Saeta's files. Saeta is a guardrail against an agent's mistakes, not containment for a hostile one.