Governed worker infrastructure

Built in conversation.
Governed in production.

Describe a job, review the exact authority, simulate it without writes, and deploy an immutable AI worker that keeps running after the conversation ends.

Build a worker in the demo View on GitHub

Open-source MVP · deterministic public demo

Worker builder · Revision 2

Inbound Sales Guardian

Ready to save

You

Require approval before any external email and cap outreach at ten per day.

AgentCloud

Updated authority and instructions. The deployment remains unchanged.

Capabilities registered9 curated

Authority explicitDefault deny

Connections ready3 fixtures

Spec hash9f3c7a2e…f10cSave immutable version

PersistentWorkers outlive the chat

Tenant-safeIsolation at every boundary

ObservableOne continuous timeline

ReversibleVersion and roll back

The operating loop

From intent to supervised work.

One persistent lifecycle across the Worker Studio and the authenticated AgentCloud MCP.

Describe

Turn an outcome into a reviewable proposal.

Simulate

Exercise the path with every write suppressed.

Govern

Review authority, budgets, readiness, and hash.

Deploy

Activate one exact immutable WorkerSpec.

Observe

Follow runs and resume human checkpoints.

Refine

Create a new version or roll back safely.

Canonical demo

Inbound sales, without the black box.

Build and refine the reference worker, inspect every capability, test the actual path without writes, deploy, pause for an exact email approval, resume, and roll back. Deterministic adapters keep the full path credential-free.

Run the guided lifecycle

worker-spec.yamlv1 · immutable

schemaVersion: "1.0"
identity:
  name: Inbound Sales Guardian
authority:
  defaultEffect: deny
  rules:
    - gmail.search_messages: allow
    - gmail.send_email: require_approval
budget:
  perRunUsd: 1

Built-in governance

Safety lives below the model.

Authority, isolation, approval integrity, dry-run suppression, and retry behavior are enforced by the control plane—not prompt suggestions.

Default-deny authority

Every integration action must be registered and explicitly granted in the WorkerSpec.

Write-safe simulation

Test the complete decision path while external side effects remain impossible.

Human checkpoints

Pause before sensitive work, inspect the exact request, then resume the same run.

Retry-safe effects

Stable idempotency keys prevent duplicate messages and records across retries.

Authenticated MCP

Your AI client can build. AgentCloud stays in control.

Start and refine persistent builder sessions from ChatGPT, Codex, or another OAuth MCP client. The AI gets safe proposal state and stable continuation links; tenant identity, deployment, policy, approvals, and credentials stay server-controlled.

Inspect the MCP implementation

01start_worker_builderworkers:write

02refine_worker_builderrevision + message

03commit_worker_builderexact hash

04test_workerzero writes

05deploy_workerseparate scope

Open the control plane

Give AI workers a durable place to work.

Launch the interactive demo Explore the source

Built in conversation.Governed in production.