Skip to content

Open source · self-hosted · AGPL-3.0

A system of record for operations, run by agents and people

Write a procedure once as a runbook. Agents and people each take their steps, and every run records who did what, when, and the evidence behind it.

Verify api after v2.4.0
post-deploy · started by the deploy webhook
Succeeded

The graph above is this runbook. An agent claimed both checks through the CLI and reported what it found. A person on call made the decision, because agents can never cast an approval. Roll back would have run only if they had chosen it. The run keeps all of it as a record you can read, search and export.

post-deploy.yaml
metadata: { id: post-deploy }
spec:
inputs:
service: { type: string, required: true }
steps:
- id: verify
parallel:
- id: check_health
actor: agent
check: true
instructions: Pass if /healthz returns 200 within 500 ms.
- id: compare_error_rate
actor: agent
check: true
instructions: Compare the 5xx rate with the last hour.
- id: decide
approval:
prompt: Keep ${{ inputs.service }} at this version?
options: [keep, rollback]
approvers: { teams: [oncall] }
- id: roll_back
if: "${{ steps.decide.outputs.decision == 'rollback' }}"
actor: agent
destructive: true
instructions: Roll back to the previous release.
The recordrun of api v2.4.0
  1. Deploy webhook

    Started the run for api v2.4.0.

  2. deploy-checkeragent

    Check health passed: 3 × 200 in 140 ms. healthz.log

  3. deploy-checkeragent

    Compare error rate passed: 0.2% against 0.3% the hour before.

  4. Sam Ortizon-call

    Approved keep: Error rate is flat, keeping it.

  5. Roll back

    Skipped, because the decision was keep. The run succeeded.

Export any run as Markdown, JSON, JUnit or PDF.

The Runbook is a system of record for operational work, such as post-deploy checks, incident triage, migrations and releases, whether an AI agent or a person does each step. You write the procedure once; each time it runs, the server tracks every step to completion and keeps who did it, what they reported and the evidence they attached.

One file per procedure

Steps, inputs, approvals, loops and parallel branches in a short YAML file you keep in git. The server checks it before anything runs.

Agents do the toil

Claude Code, Codex, Cursor or your own agent claim steps under a lease, report outputs and attach evidence, through the CLI or MCP.

People keep the judgment

Approvals by named people, destructive steps that need a human, take-overs, reminders and escalation. Agents never cast human approvals.

A record you can read

Every run reads like a document: who did each step, their outputs, logs and evidence. Export it as Markdown, JSON, JUnit or PDF.