Skills

Install with skills CLI

A collection of skills for coding agents that work with IndyKite — graph-based identity, authorization, and AI-agent integration.

A skill is a self-contained bundle of instructions (and optionally scripts, references, or assets) that an agent loads on demand to perform a specialized task.

Contents

Glossary

Acronyms used throughout the skills, defined once here.

  • IndyKite — the company behind these skills. UI at eu.hub.indykite.com / us.hub.indykite.com; REST API at openapi.indykite.com; docs at developer.indykite.com.
  • IKG — IndyKite Graph; the graph database holding identity, relationships, and properties for a project.
  • CIQ — ContX IQ; an IndyKite context authorization policy plus its Knowledge Query, executed via POST /contx-iq/v1/execute.
  • Knowledge Query — the parameterised read/write definition CIQ runs against the IKG.
  • AuthZEN — IndyKite's policy-decision endpoint (authzen_evaluate).
  • IAG — IndyKite Agent Gateway; enforces caller, workflow, and delegation-chain checks in front of A2A agents or MCP servers.
  • A2A — agent-to-agent; one autonomous agent calling another.
  • MCP — Model Context Protocol; the IndyKite MCP server exposes AuthZEN and CIQ tools to MCP-aware agents.

Skills in this repo

Each row is one skill — what it does and a representative prompt it's designed to handle. Activation isn't guaranteed: which skill actually fires depends on the agent's matching algorithm, the model, and what else is installed. Verify routing in your own setup before relying on it.

Skill What it does Example prompt
indykite-agent-gateway Deploy and configure IAG in front of an A2A agent or MCP server to enforce caller, workflow, and delegation-chain checks. "Put IAG in front of my MCP server so every call is authorized before it reaches it, and wire up the workflow in the IKG."
indykite-mcp-server Make AuthZEN/KBAC authorization decisions and run ContX IQ Knowledge Queries from an AI agent over the Model Context Protocol — initialize a session, list tools, call the AuthZEN and CIQ tools — and configure the MCP endpoint for a project. "How do I initialise an MCP session against eu.mcp.indykite.com and call authzen_evaluate?"
indykite-capture-upsert-nodes Build the request body for the Capture API batch node upsert (POST /capture/v1/nodes) — entities with properties, metadata, and external-data references. "Add these people and cars to the IKG, with verified email metadata."
indykite-capture-upsert-relationships Build the request body for the Capture API batch relationship upsert (POST /capture/v1/relationships) — typed connections between existing nodes. "Link millicent OWNS kitt and ryan CAN_DRIVE kitt."
indykite-capture-delete-nodes Build the request body for the Capture API batch node delete (POST /capture/v1/nodes/delete) — remove whole nodes by external_id + type. "Delete the test Person and Car nodes from the graph."
indykite-capture-delete-node-properties Build the request body for the Capture API node-property delete (POST /capture/v1/nodes/properties/delete) — strip named properties, keep the node. "Drop the email property from millicent."
indykite-capture-delete-node-property-metadata Build the request body for the Capture API property-metadata delete (POST /capture/v1/nodes/properties/metadata/delete) — strip provenance fields, keep the value. "Remove the assurance_level metadata from millicent's name property."
indykite-capture-delete-relationships Build the request body for the Capture API batch relationship delete (POST /capture/v1/relationships/delete) — disconnect nodes, keep them. "Remove the CAN_DRIVE link between ryan and kitt."
indykite-capture-delete-relationship-properties Build the request body for the Capture API relationship-property delete (POST /capture/v1/relationships/properties/delete) — strip properties from an edge, keep the edge. "Drop the status property from millicent's OWNS edge."
indykite-authzen-kbac-policies Author and manage a KBAC authorization policy (subject, actions, resource, Cypher condition) through the Config API — create, read, list ?type=kbac, update, delete. The home of KBAC policy authoring. "Write a policy letting a Person PROVISION a Server when its cost is within a budget, and publish it."
indykite-authzen-evaluation Make a single KBAC decision via the AuthZEN REST API (POST /access/v1/evaluation) — one yes/no answer for a (subject, action, resource) triple. "Can ada PROVISION the server gpu-node-7 with a budget of 120000?"
indykite-authzen-evaluations Run many KBAC decisions in one call (POST /access/v1/evaluations) with top-level defaults overridden per entry. "Of these servers, which can grace provision with a budget of 80000?"
indykite-authzen-search-action List the actions a subject may perform on a resource (POST /access/v1/search/action). "What can linus do with the server gpu-node-7?"
indykite-authzen-search-resource List the resources a subject may act on, given an action (POST /access/v1/search/resource). "Which servers can linus provision with a budget of 4000?"
indykite-authzen-search-subject List the subjects allowed an action on a resource (POST /access/v1/search/subject). "Who can provision the server gpu-node-7 with a budget of 80000?"
indykite-ciq-read Author a read-only CIQ policy and Knowledge Query, then run it via POST /contx-iq/v1/execute. "Expose Person-OWNS-Car as a parameterised read query through ContX IQ."
indykite-ciq-create-node Author a CIQ policy + Knowledge Query that creates a brand-new node in the IKG. "Create a new Track node in the IKG with title and loudness, via CIQ."
indykite-ciq-create-relationship Author a CIQ policy + Knowledge Query that creates a brand-new relationship between two existing nodes. "Add a PLAYED_AT relationship between an existing Track and an existing Venue."
indykite-ciq-create-node-with-link Author a CIQ policy + Knowledge Query that creates a new node AND links it to one or more existing nodes in a single atomic execute. "Create a new Contract and atomically link it to an existing Vehicle and Person."
indykite-ciq-add-property Author a CIQ policy + Knowledge Query that sets one or more properties on an existing node. "Let a Person update their own music_mood property."
indykite-ciq-add-relationship-property Author a CIQ policy + Knowledge Query that sets one or more properties on an existing relationship. "Annotate an existing PLAYED_AT relationship with a verified flag and timestamp."
indykite-ciq-delete Author a CIQ policy + Knowledge Query that deletes a node, a relationship, or one or more properties. "Clear the music_mood property from a Person — GDPR-style erase."

Install

The recommended path is the skills CLI — one command, many agents. The CLI knows the per-agent install location and targets one (or all) via the --agent flag.

# All skills in this repo, into the project's local agent directories
npx skills add indykite/skills

# All skills, globally for the current user
npx skills add indykite/skills -g

# Just one skill, into one agent
npx skills add indykite/skills --skill indykite-mcp-server --agent claude-code

# Show what's in the repo without installing
npx skills add indykite/skills --list

Useful flags:

  • -g, --global — install at user scope instead of project scope.
  • -a, --agent <name> — limit to one or more agents (use * for all). A typo prints the full valid set.
  • -s, --skill <name> — limit to one or more skills.
  • --copy — copy files instead of symlinking (default is symlink so updates propagate).
  • -y, --yes — skip confirmation prompts (handy in CI).

After installing, restart the agent (or reload its config) so it picks up the new skill directory. Verify with npx skills list.

Quick verification: install one skill, then ask the agent something its description matches. For indykite-mcp-server: anything about initialising an MCP session against eu.mcp.indykite.com / us.mcp.indykite.com, calling authzen_evaluate / ciq_execute, or debugging a 401 from the MCP server.

Manual install

If you don't want the CLI, copy or symlink the skill directory into the agent's expected location. For Claude Code that's ~/.claude/skills/<skill-name>/ (user scope) or .claude/skills/<skill-name>/ (project scope). For any other agent, consult its docs — most that don't read SKILL.md natively expect the body (everything after the YAML frontmatter) pasted into their own rule format.

Bundle install

For agents that support a single labelled plugin install, this repo ships per-agent manifest files. They register every skill in this repo at once and prompt for credentials (API_URL, API_KEY, BEARER_TOKEN, SERVICE_ACCOUNT_TOKEN, MCP_URL, PROJECT_GID) at install time. Source of truth for the credential list is manifest.yaml.

Claude Code

Ships .claude-plugin/plugin.json and .claude-plugin/marketplace.json:

# 1. Register this repo as a marketplace
/plugin marketplace add indykite/skills

# 2. Install the plugin from it
/plugin install indykite-skills

Gemini CLI

Ships gemini-extension.json at the repo root:

gemini extensions install https://github.com/indykite/skills

Gemini uses this README.md as its loaded context (contextFileName) and exposes each settings entry as an environment variable.

Cherry-picking from a bundle

If you only want one or two skills, use npx skills add indykite/skills --skill <name> instead — see Install. The bundle prompts for all six credentials; per-skill installs leave credential management to env vars.

Supported agents

The skills CLI is a cross-agent installer — npx skills add indykite/skills --agent <name> drops files in the right place for whatever agent <name> the CLI knows about. The CLI's --agent list is the source of truth and stays current as new agents are added.

Verified end-to-end in this repo: Claude Code, Gemini CLI.

For every other agent the CLI lists, files arrive at the right path — but native SKILL.md triggering is a property of the agent, not the CLI. Whether an agent activates a skill automatically (matching the prompt against description), via a slash command, or by explicit selection depends on that agent. Check its docs when automatic activation matters.

How skills activate

A skill is passive until invoked. Supported agents activate skills automatically by matching the user's prompt against each installed skill's description field — the one-line summary in SKILL.md's frontmatter. When the description fits the request, the agent loads the rest of the skill into context before answering. The ## When to use section is also read by the agent and determines whether activation fires — that's why the skills here list both positive triggers and explicit anti-triggers.

If a skill doesn't activate when you expect it to, check three things:

  1. The agent loaded it. npx skills list shows what's installed for the current scope. A skill that isn't installed cannot activate.
  2. The description fits the prompt. Open SKILL.md and re-read description and ## When to use. Vague descriptions get vague triggering — sharpen the wording, reinstall, retry. If two skills could match the same prompt, the agent picks one; tighten one description to disambiguate.
  3. Manual fallback. Most agents let you invoke a skill by name (/<skill-name> in Claude Code, or explicit selection in Cursor / Copilot). Useful when automatic routing is uncertain — also tells you whether the skill is loaded at all.

To disable a misbehaving skill: remove the directory (npx skills remove <name>, or delete from the agent's skills folder), or set metadata.internal: true in its frontmatter so it stays installed but hidden from automatic discovery (it surfaces only when INSTALL_INTERNAL_SKILLS=1).

Authoring a skill

At-a-glance reference; full submission flow lives in CONTRIBUTING.md.

Structure

/
├── skill-name/
│   ├── SKILL.md
│   ├── scripts/      (optional)
│   ├── references/   (optional)
│   └── assets/       (optional)
└── another-skill/
    └── SKILL.md

SKILL.md format

---
name: your-skill-name
description: What it does. Use when [trigger conditions].
---

# Your Skill Name

## Instructions

### Step 1: ...
### Step 2: ...

Conventions

  • Folder names are kebab-case (e.g. my-cool-skill).
  • The file must be named exactly SKILL.md (case-sensitive).
  • The description should state both what the skill does and when to invoke it — agents use it to decide whether the skill is relevant.
  • Keep SKILL.md focused on core instructions; put longer docs in references/ and helper code in scripts/.

Resources

Contributing and security

  • Add a new skill, fix one, or change conventions → CONTRIBUTING.md covers the quality bar, style, testing checklist, Agent Skills specification compliance, submission process, and code of conduct.
  • Exercise the skills (structural validation, dry-run smoke tests, live API roundtrips) → testing/README.md. Harness: ./testing/e2e-ciq.sh.
  • Found a security issue → responsible_disclosure.md before opening a public issue.
  • The LICENSE at the repo root applies to every file in the tree.