Approvals and Auto
Who decides what a thread may do, in order, and how Auto's AI reviewer fits in
Every tool call a thread makes is checked before it runs. Reading code and editing files inside the thread's worktree run without asking. Everything else is decided in this order, and the first step that decides wins:
- Your limits in Settings → Agents. Nothing below can allow what your limits ask about.
- The major's rules, for threads in a major's crew (Majors).
- The thread's permission mode: Ask before running, or Auto, where an AI reviewer judges the request.
- You.
Your limits
Settings → Agents lists what threads may do, each set to Allow or Ask. A change applies to the next tool call, in every thread and major.
| Action | Default |
|---|---|
| Read code and run tests | Allow |
| Install packages already in the lockfile | Allow |
| Run the dev server and open the browser | Allow |
| Push to branches the thread created | Allow |
| Open draft pull requests | Allow |
| Use a connected service | Allow |
| Migrations, and anything else that changes data | Ask |
| Push to main, and merge | Ask |
| Spend money: paid APIs and cloud machines | Ask |
| Stop processes it didn't start, or control other apps | Ask |
| Use a secret | Ask |
A few actions always ask, whatever your limits say: writing or deleting outside the worktree, force-pushing, and resetting a branch the thread didn't create.
Hard rules
Some actions ask in every mode, including Auto, a major's rules and Always allow:
- Stopping, quitting or restarting Keko itself: the app, its engine or the
kekoservice. - Touching Keko's own data folders or its keychain item.
- Deploying or changing production, such as
vercel --prod,fly deploy,kubectl applyon a remote cluster, or a release workflow.
The hard rules guard against accidents and against an agent that's confused or was given bad instructions. When a command is too complex to read, they ask rather than guess.
Answering a request
A request shows as a card above the composer with the command and what it does: Don't allow · Allow ⌘↵. What you type in the composer goes back with Don't allow, so the thread knows what to do instead. Requests also show in the sidebar row, in a major's chat, in notifications and on the iPhone, and answering in one place clears the others.
For kinds of actions no limit covers, such as adding a package, opening a pull request or using the network, the card offers Always allow for the rest of the thread.
Requests wait durably: if the app quits or the engine restarts, the request is still there when it comes back, and the call runs once you answer.
Auto
With Auto, each request that would ask you is first judged by a model, from the thread's conversation, your limits and the exact call. Choose that model in Settings → Agents → Reviewer model.
| The reviewer finds | What happens |
|---|---|
| Low or medium risk, and you asked for it | It runs, and the thread logs "Auto allowed: …" |
| High risk, or not something you asked for | It asks you, with the reviewer's reason |
| Critical | It's blocked and logged, and the agent is told why |
A review that fails or takes too long asks you. Your limits still come first: Auto never allows something your limits ask about. Settings → Agents also sets whether new threads start in Ask before running or Auto.
A cloud machine's desktop
Threads on Keko Cloud machines can use the machine's desktop (Proof of work). Looking at the screen never asks. The first click or keystroke asks once, as Use the machine's desktop, and allowing it allows the desktop for the rest of the thread. Opening an address follows the same limits as opening it from a command: localhost is "Run the dev server and open the browser", anything else "Use the network".
Plan mode
/plan asks the thread for a plan before it changes anything. Until you approve, edit or reject the plan, the thread
can only read and run read-only commands.