Give your agent a Worker, not your keys.
A Worker is a scoped, audited identity for an AI agent. It can read your infrastructure and propose changes you approve - and it can never delete resources or read your secrets. Run it unattended on the model key you bring, and every action is logged.
Free to start, no credit card · Bring your own model key
One identity. Reads flow, changes wait for you, secrets stay off limits.
Three things can happen. You control which.
Handing an agent your API key gives it the power to delete everything and read every secret. A Worker splits every possible action into three lanes - and only one of them can act without you.
How a Worker run works.
From "here's the task" to "done, and logged" - with you in the loop for anything that matters. Click a step to see what happens.
Pick how far it can go.
Two roles, set per Worker, changeable any time. Start read-only and graduate to proposing changes.
Observer
Reads, and nothing else.
- ✓Read logs, metrics & deployments
- ✓Browse your connected source
- ✓Summarise and diagnose
- ✕Propose or make any change
- ✕Touch secrets or env vars
Maintainer
Reads, and proposes changes you approve.
- ✓Everything an Observer can
- ✓Work in isolated sandboxes
- ✓Propose deploys, rollbacks & pull requests
- ✕Run a change without your approval
- ✕Delete resources or read secrets
Make it yours
Shape how each Worker works.
Roles decide what a Worker may do. These decide how it does it - its identity, its playbooks, and how big a job it can take on. None of them widen what its role allows.
Standing instructions
Tell a Worker who it is and how you work - "you are our release manager; always check production logs before proposing a deploy." It follows that on every run, on top of each task.
Skills
Hand a Worker reusable playbooks. Write a review or triage skill once, name it in a task, and the Worker loads the right one and follows your steps instead of guessing.
Swarm
For a big job, let one Worker fan the work out to parallel helpers and gather the results - a single Worker that scales into a team for one task, still under one approval queue.
Native to your whole 
orkestr account.
A Worker isn't a bolt-on bot. It already understands your projects, functions, environments, crons and jobs - so it can act in context, not guess. It sees exactly what its role allows, and nothing more.
Projects & deployments
Live status, deploy history, and rollouts across every project it is scoped to.
Serverless functions
Your functions, their logs and metrics - and, for a Maintainer, updates it can propose.
Environments & domains
Production, staging and preview environments, their branches, config and URLs.
Crons & jobs
Scheduled runs and one-off jobs - their timing, status, exit codes and output.
Logs & metrics
Container logs, CPU and memory, request rates - the signals it needs to diagnose.
Your source code
Browses the connected repo to reason about the real change, not just the symptom.
What people hand to a Worker.
Anywhere you'd want an agent to act on your infrastructure - without giving it the keys to wreck it.
Incident triage on autopilot
An error spike wakes a Worker. It reads the logs and metrics, finds the regression, and proposes a rollback - waiting for your nod.
Ship fixes from chat
Point your coding agent at a Worker. It reads the code, prepares the change, and opens it for approval - without ever holding your real key.
Give a vendor scoped access
Hand a contractor or third-party agent a Worker instead of credentials. They can read and propose; they can never delete or exfiltrate.
Unattended, auditable automation
Recurring checks and cleanups that run on their own model, pause at the gate for anything risky, and leave a full paper trail.
A team for your project
Run a team of specialists.
Give one project several Workers - each owning a slice. They do not step on each other, and everything that changes your infrastructure lands in the same approval queue: yours.
Release Manager
Reacts to a failed deploy
Pulls the build and runtime logs, diagnoses the likely cause, and proposes a rollback or fix - for your approval.
Code Reviewer
Reacts to a new pull request
Reads the diff the moment a PR opens and posts a review of what changed and anything risky it spots.
Morning Reporter
Runs on a schedule
Each morning it checks deploys, health, and usage and posts a status digest to your Slack.
Built for the day the agent is wrong.
Every Worker assumes its agent can be manipulated by the data it reads. So the boundary does not depend on the agent behaving. Destructive actions are unreachable, secrets are off limits, and anything irreversible waits for a human. You delegate the work without surrendering control, and you keep a complete record of everything that happened.
Common questions
What can a Worker actually do?
What if the agent is tricked or goes wrong?
Which agent or model do I use?
How is this different from just making a limited API key?
Where does it run, and who can see it?
Stop handing agents your keys.
Create a Worker, hand it to your agent, and let it work - safely.
Start free