Agent Skills

Let your coding agent control Kumo through a small, safe, well-described command surface.

Modern coding agents (Cursor, Claude Code, Codex, Gemini Code Assist, Agents.md-style configs) read skill files from well-known directories so they know what they can do in the environment.

Kumo ships a bundled skill — kumo-cli — that teaches an agent:

The JSON envelope shape (ok, data, error).
The safe set of commands the agent should use.
The --dry-run convention for anything that writes system state.
Where to look for logs when something goes wrong.

The skill restricts agents to read-only state queries, node switching, subscription refreshing, and log inspection. Arbitrary shell execution is not allowed.

Install the skill

From the GUI

The first-launch onboarding sheet asks which agents to enable. This can also be run later from Settings → General → Agent Skills, or from the Agent Skills destination in the sidebar.

For each supported agent, pick:

Global — install into the agent's user-wide skill directory.
Project — install into the current project (limited to agents that support per-project skills).

Each row shows the destination path, whether the skill is currently installed, and whether it is up to date with the version inside the running Kumo.

From the CLI

kumo skills status --json
kumo skills install --agent claude --scope global
kumo skills install --agent all
kumo skills install --agent cursor --scope project --dry-run --json
kumo skills uninstall --agent cursor --scope project

--agent accepts cursor, claude, codex, gemini, agents, or all. --agent all --scope project only targets agents that support project scope.

Skill paths

Agent	Global path	Project path
Cursor	`~/.cursor/agents/kumo-cli`	`./.cursor/agents/kumo-cli`
Claude Code	`~/.claude/skills/kumo-cli`	`./.claude/skills/kumo-cli`
Codex	`~/.codex/skills/kumo-cli`	— (not supported)
Gemini	`~/.gemini/skills/kumo-cli`	— (not supported)
agents.md	`~/agents/kumo-cli`	`./agents/kumo-cli`

If a skill directory exists at the destination but Kumo did not write it (for example a colleague's setup was cloned), install fails with a clear error. Pass --force to overwrite — Kumo first moves the existing directory aside.

Uninstall is similarly conservative: only entries Kumo recorded as its own are removed. Anything else stays.

Re-running install after a Kumo update brings the skill content up to date. Until then, the agent will still work — it just will not know about new commands added in the newer version.

The skill itself

The skill is a small Markdown bundle that teaches the agent to call kumo for everything. The agent does not need any Kumo-specific code in its prompt — the skill is the contract.

Read the bundle that ships with the current Kumo by opening:

open "$(kumo skills status --json | jq -r '.targets[0].destinationRoot')"

(Or browse to the relevant path above for the agent.)

Cleanup

kumo skills uninstall --agent all

Removes every skill Kumo wrote. Untracked entries — for example a hand-edited copy — are left in place.