Skip to content
CommandChain CommandChain CommandChain Docs

Skill Graph Studio

Settings → AI Configuration → Skill Graphs is where the skills your agents can draw on are defined, checked, and released. It needs the manage_skill_graphs permission.

A skill graph is your workspace’s answer to “what does my team know how to do”. Agents resolve skills from the published graph, so editing here is safe: nothing an agent does changes until you publish.

Draft and published

Section titled “Draft and published”

There are always two graphs.

The draft is what the studio edits. The published graph is what agents resolve against. They are separate on purpose: a half-finished skill in the draft is invisible to every run.

The three tabs are Skills, Edges, and Publish.

Skills

Section titled “Skills”

A skill is one capability. Each carries:

  • Kind: atomic, or composite (built from others).
  • Capabilities, contexts allowed, and agent types allowed: where it may be used.
  • Effects: what it changes in the world.
  • Risk tier and governance: whether using it needs approval, which scopes it needs, and which budget class it draws on.
  • Guidance: the prose an agent actually reads.

Guidance is the part worth spending time on. A skill with a perfect risk tier and no guidance tells an agent that a capability exists and nothing about when to reach for it.

Bindings attach a skill to a real tool: an MCP tool, a local tool, a connector, or an executable skill. A skill with no binding is a description of a capability that cannot be exercised.

Edges

Section titled “Edges”

An edge relates two skills: one requires another, one refines another, one is an alternative to another. Edges are what let an agent move from the skill it has to the skill it needs.

Validate

Section titled “Validate”

Validate checks the draft and answers with errors and warnings. Errors block publishing: a binding pointing at a skill that no longer exists, an edge whose endpoint is missing. Warnings do not, but they are worth reading, because most of them are the shape of a skill nobody can use.

Publish

Section titled “Publish”

Publishing takes a snapshot of the draft, gives it a version, and makes it the graph agents resolve against. Add a changelog line saying what changed and why; it is what the version list shows later.

Validation runs before publishing, so a graph with errors cannot become the live one.

Roll back

Section titled “Roll back”

Version History lists every published version. Rollback makes a previous version live again.

Roll back when a published graph makes agents worse, and figure out why afterwards. That is a faster loop than editing a live graph under pressure.

Next steps

Section titled “Next steps”