Agent Playbook OS, open source
Run AI work from a plan you can check.
You write the steps in a YAML playbook. Agent Playbook OS runs them in order, stops where it needs your approval, picks up where it left off and keeps a record of what happened.
git clone https://github.com/CometWeb-io/agent-playbook-os.git Source preview, not on PyPI Licence MIT Python 3.11+, playbook CLI
What it does
Plan first. Then run.
Before anything runs, the playbook is validated and compiled into a fixed plan. You know which steps will run, in what order and where the run will stop.
playbook compile
The order is set before the start.
Compiling checks the schema, the dependencies between steps and the policies, then saves the plan with its hash. You run exactly the plan you reviewed.
- Schema and dependency checks
- A plan with a semantic_hash
- Policies outrank model instructions
gate: human
Stop where a person decides.
A gate step pauses the run in WAITING_APPROVAL. After approval you resume from that point, and the record keeps who approved it.
- Approval as an ordinary playbook step
- Resume without repeating earlier steps
- The approver named in the record
playbook trace
A record you can come back to.
Every run event goes into a SHA-256 chain. The trace, replay and diff commands show what happened and how two runs differed.
- A contiguous SHA-256 event chain
- Inspect, replay and compare runs
- Secrets stored as references, not values
How it works
From clone to first run.
- Install from a clone Python 3.11+, a virtual environment and pip install -e. It is not on PyPI yet.
- Check the playbook playbook validate and playbook compile show the plan before anything runs.
- Run and approve The example run stops at the approval step. playbook resume finishes it.
- Read the record playbook trace lists the run’s events in order.
Installation
Run it from a clone.
Until the first release, install Agent Playbook OS from a copy of the repository in a Python virtual environment. The kernel-demo example stops for human approval, so you see resuming work straight away.
- Python 3.11 or newer
- A virtual environment and pip
- Git to clone the repository
git clone https://github.com/CometWeb-io/agent-playbook-os.git
cd agent-playbook-os
python3 -m venv .venv && source .venv/bin/activate
python -m pip install -e ".[dev]"
playbook run playbooks/examples/kernel-demo.yaml --input name=Ada --run-root .playbook-runs Finish the run after approval.
playbook resume .playbook-runs/<RUN_ID> --approve approval --approval-actor you@example.comLimits
This is an early preview.
Interfaces may still change between versions. The Agent Skills integration and host comparisons are structural and do not confirm live support for a specific Claude, Codex or ChatGPT host.
- No PyPI release yet, install from a clone
- AI host integrations are not verified live yet
- Capability declarations are not a security certification
Code and docs
Everything lives in the repository.
No. One operator runs the playbook, and separate agents run only where isolation helps, for example for an independent review.
No. The runtime and state storage are defined by provider-neutral contracts. The kernel-demo example runs without a model.
An Agent Skill describes how to do one task. A playbook sets the order of steps, the approvals and the record of the whole run. They can work together.
From website analysis
to action.
CometWeb Insight checks five quality areas, prioritises fixes and re-tests the site after you ship.