delego — Claude Code plugin

The one-command way to get delego — intent-bound action authorization for AI agents — into a Claude Code project: its skills, review agents, and the MCP server, all in a single plugin. This repo is both the plugin and its own marketplace.

Install

In Claude Code:

/plugin marketplace add Delego-Dev/plugin
/plugin install delego@delego

That's it — Claude Code places the skills, agents, and MCP server in the right spots (no manual cloning into .claude/). Then:

/delego:init

/delego:init installs the delego Python package (pip install "delego[mcp]"), creates the project's delego home at .claude/.delego, drafts/validates a policy, and verifies. The plugin's MCP server reads DELEGO_HOME = <project>/.claude/.delego and activates once the package is installed and the home exists.

Prerequisite: the MCP server runs the delego-mcp command, which comes from pip install "delego[mcp]" (done by /delego:init). Until it's installed, Claude Code's /plugin Errors tab will show delego-mcp not found — expected; it clears after init + a reload.

What's inside

Skills (/delego:<name>)

Skill	Use it to…
init	Install delego, create the delego home, verify.
policy-drafter	Draft/harden a fail-closed policy.yaml.
approval-triage	Review pending approvals and approve/deny.
audit-explainer	Verify the signed chain and explain what the agent did.

Agents (@delego:<name>)

Agent	Use it to…
policy-reviewer	Adversarially review a policy.yaml (footguns, over-broad rules, fail-open gaps).
broker-reviewer	Review a custom BrokerAdapter against the invariants.
audit-investigator	Verify the ledger, reconstruct authority paths, flag anomalies.

MCP server — delego_propose_action, delego_resolve_action, delego_audit_tail, delego_show_policy (the agent's interface to delego).

Typical flow

1. /plugin install delego@delego → /delego:init (sets everything up).
2. /delego:policy-drafter to write the policy → @delego:policy-reviewer to harden it.
3. At runtime the agent proposes actions through the MCP server; decide parked ones with /delego:approval-triage, read the trail with /delego:audit-explainer, investigate incidents with @delego:audit-investigator.

Repo layout

.claude-plugin/marketplace.json    # this repo is its own marketplace
delego/
  .claude-plugin/plugin.json       # the plugin manifest
  .mcp.json                        # the delego MCP server
  skills/<name>/SKILL.md           # invoked as /delego:<name>
  agents/<cluster>/<name>.md       # invoked as @delego:<name>

python scripts/validate.py (run in CI) checks the manifests and every skill/agent's frontmatter.

Without the plugin (à la carte)

Prefer not to use a plugin? Clone this repo and run install.sh from your project root — it copies the skills and agents into .claude/ with a delego- name prefix (so /delego-init, @agent-delego-policy-reviewer):

git clone https://github.com/Delego-Dev/plugin /tmp/delego-plugin
cd your-project && /tmp/delego-plugin/install.sh

(The standalone Delego-Dev/skills and Delego-Dev/agents repos are archived — this repo is the single source of truth.)

License

Apache-2.0. Built for delego; see the wire specification.