Central Agentic Ops
[!NOTE]
Research prototype: Features and interfaces may change.
[!TIP]
Agents: Use the central-agentic-ops skill
to set up a control plane, operate campaigns, analyze evidence, or debug CAO.
Central Agentic Ops (CAO) brings GitHub Agentic Workflow runs, costs, and
outcomes into one dashboard, then adds policy-governed execution when you are
ready. It operates from a control repository across explicitly enrolled
repositories. CAO does not replace repository permissions, branch protection,
or human review.
Open a local dashboard
Use your locally authenticated GitHub CLI to install
the release extension and inspect a repository:
gh auth status
gh extension install githubnext/gh-cao
gh cao dashboard --repo OWNER/REPOSITORY
No CAO cloud deployment, Azure setup, hosted dashboard, Redis, or separate
Node.js or gh-aw installation is required. A running Docker or Podman engine
provides PostgreSQL automatically unless a connection is specified. From an
existing repository checkout, use gh cao dashboard without --repo.
No gh cao setup or control-repository configuration is needed.
Open the Dashboard: URL printed in the terminal. The bundled Go server
serves on loopback and queries local PostgreSQL. Managed data is bind-mounted
under the ignored .github/cao/dashboard/<scope>/postgres/ cache.
Docker is tried first, then Podman. Pass --container-engine docker|podman (or set
CAO_CONTAINER_ENGINE) to use only that engine. --postgres-url or
CAO_POSTGRES_URL bypasses container management and uses provisioned storage.
Collection uses your GitHub CLI identity to read GitHub evidence and refreshes
automatically every five minutes: no cloud deployment does not mean offline.
The recent evidence window and session lifetime default to, and cannot exceed,
48 hours; collection is bounded, not a complete historical backfill.
Press Ctrl-C to stop serving and collection while retaining the cache.
See the local dashboard quickstart
for scope, cache reuse, and evidence limits.
Set up shared monitoring when ready
Git and authenticated GitHub CLI 2.68+ are required. Need a repository?
Create one with gh repo create OWNER/REPOSITORY --private --clone, then
cd REPOSITORY.
From the repository root:
gh cao setup
Setup makes this the control repository by preparing the monitoring runtime,
policy, and gh-aw. It does not install or run a campaign, and it keeps
changes local unless you separately approve commit and push. A published setup
reports deployment pending; an empty dashboard is normal until Agentic
Workflow evidence exists.
Use a private control repository for non-public evidence. Production dashboards
use same-origin Go/PostgreSQL hosting; local development uses the same Go server
over a verified canonical JSONL artifact ingested into PostgreSQL. GitHub Pages hosts only the Synthetic enterprise
demo, with the same canonical pipeline, queries, and views but no live services.
Setup does not configure Pages.
[!TIP]
SAML-token installation failure: Use the
curl installer:
curl -fsSL https://raw.githubusercontent.com/githubnext/gh-cao/main/install-gh-cao.sh | bash -s -- --setup.
Follow Set Up CAO for authentication, scope,
publication, and your first dashboard visit.
Add work when ready
Start with one repository in review; live outputs require explicit reviewed
policy and declared safe-output capabilities. Describe the outcome and scope
first—campaigns are one way to scale work, not a prerequisite. See
From Outcome to Execution and the
campaign catalog.
You can also inspect verified Repo Assist examples
without running more work.
For agents, install the CAO plugin and use the
central-agentic-ops skill, which routes each task to one guide under
.github/cao/instructions/.
Read the docs
Want to contribute? Read Contributing and follow the
Code of Conduct. For help, check Support.
Report vulnerabilities privately through Security.
Maintained by GitHub Next.
Acknowledgements: CAO's service design principles are inspired by SAP's
Calesi pattern for CAP-Level Service Integration:
domain-level interfaces, protocol-independent consumption, fast local development,
interchangeable implementations, and deliberate extensibility.
Licensed under the MIT License.