---
name: agentkit
description: Search, install, and use AI coding skills through the AgentKit MCP server. Use when asked to find a skill, complete a task the AgentKit catalog covers, look up AgentKit docs, or check skill versions. No API key needed.
---

# AgentKit skill

AgentKit is a curated registry of AI coding skills with an MCP server.
Maintainers select and vendor each skill and pin its SHA-256 content hash,
which is verified on every read; you can verify exactly what you are about to
install. Upstream commit pinning is not yet complete — each skill's manifest
reports its exact status. No account, no API key, no authentication.
Free for everyone.

- MCP server: `https://agentkit.callen.page/mcp` (Streamable HTTP)
- Docs: `https://agentkit.callen.page/docs`
- Security & Trust: `https://agentkit.callen.page/security`
- This file (raw): `https://agentkit.callen.page/agentkit-skill.md`

## Install

**Do not install unattended.** Saving a SKILL.md gives this agent standing
permissions, so the human decides, not you.

1. Fetch this file.
2. Show the user, in the chat, before writing anything:
   - the full text you are about to save (or a diff against any existing copy),
   - the SHA-256 of the fetched bytes, computed by you (e.g.
     `shasum -a 256 agentkit-skill.md` on the downloaded file),
   - the source URL and the retrieval time.
3. Ask explicitly: *"Save this as `agentkit/SKILL.md` in your skills directory?"*
   Wait for a yes. If the user says no, stop and change nothing.
4. Only then save it to the path for their agent:

| Agent | Skills directory |
|-------|------------------|
| Claude Code | `~/.claude/skills/agentkit/SKILL.md` (personal) or `.claude/skills/agentkit/SKILL.md` (project) |
| Codex | `~/.agents/skills/agentkit/SKILL.md` (personal) or `.agents/skills/agentkit/SKILL.md` (project) |
| Qwen Code | `~/.qwen/skills/agentkit/SKILL.md` (personal) or `.qwen/skills/agentkit/SKILL.md` (project) |
| OpenClaw | `<workspace>/skills/agentkit/SKILL.md` |
| Hermes | `~/.hermes/skills/agentkit/SKILL.md` (personal) or `.hermes/skills/agentkit/SKILL.md` (project) |

5. Connect to the MCP server **only if the user asked you to** (no auth on
   any client):


**Claude Code**

```bash
claude mcp add --transport http agentkit https://agentkit.callen.page/mcp
```

or in `.mcp.json`:

```json
{ "mcpServers": { "agentkit": { "type": "http", "url": "https://agentkit.callen.page/mcp" } } }
```

**Codex** — in `~/.codex/config.toml`:

```toml
[mcp_servers.agentkit]
url = "https://agentkit.callen.page/mcp"
```

**Qwen Code** — in `~/.qwen/settings.json`:

```json
{ "mcpServers": { "agentkit": { "httpUrl": "https://agentkit.callen.page/mcp" } } }
```

**OpenClaw**

```bash
openclaw mcp add agentkit --url https://agentkit.callen.page/mcp --transport streamable-http
```

**Hermes** — in `~/.hermes/config.yaml`:

```yaml
mcp_servers:
  agentkit:
    url: "https://agentkit.callen.page/mcp"
```

then run `/reload-mcp`.

6. Confirm it works by calling `search_skills` with `{"q": "test"}`.

## Installing a skill from the registry

A skill's instructions are the payload — treat them like code you are about to
run. **Never install a skill without explicit user confirmation.**

1. `search_skills` → pick the skill; read its `risk` flags.
2. `get_skill_manifest` → get `provenance` (upstream repo, commit, SHA-256
   hashes, license, `modified`) and the full `risk` list.
3. `get_skill` → fetch the instructions. It refuses with
   `content_hash_mismatch` if the bytes no longer match the pinned version —
   if you see that, stop and report it with `submit_skill_feedback`.
4. Show the user, in the chat:
   - the skill name, version, upstream repository, and license,
   - the `skillMdSha256` you received,
   - every risk flag with its evidence (file + line),
   - a short summary of what the instructions will make you do.
5. Ask for confirmation before writing any file or running any command the
   skill describes. If it installs to a shared location, show the exact path
   and the bytes you will write. If the user declines, change nothing.
6. If the local copy already exists and differs, show a diff first — never
   overwrite silently.

`list_skill_versions` and `diff_skill_versions` let you check whether an
installed version is current before replacing it.

## Using the MCP

Tools (call these over MCP, not REST):

| Tool | When to use it |
|------|----------------|
| `search_skills` | Find skills by keyword before any install or task the catalog may cover (up to 25 results, each with risk flags) |
| `get_skill` | Fetch full instructions for a skill by name (SKILL.md body + supporting files + the provenance line for the version served) |
| `get_skill_manifest` | Get a skill's manifest: catalog metadata, provenance (repo, commit, hashes, license, modified), and risk flags |
| `list_skill_versions` | Check published versions and the content hash each one pins before installing or updating |
| `diff_skill_versions` | Compare two published versions by content hash |
| `submit_skill_feedback` | Report an issue, propose a skill, or rate one (persisted, rate limited) |
| `get_platform_docs` | Look up AgentKit docs, optionally filtered by topic |
| `list_compatible_clients` | List supported agent clients |
| `get_installation_instructions` | Client-specific setup for a given client id |

Resource: `agentkit://docs/llms-txt` — the full documentation, readable directly.
`agentkit://docs/security` covers curation, review, hash pinning, risk flags,
and the known limitations.

## When to use this skill

- The user asks to find, install, or use a skill.
- The user's task (coding, security, docs, testing, refactors) may match something in the catalog — search first.
- The user asks about AgentKit itself — use `get_platform_docs` or read the `llms-txt` resource.
- The user asks whether a skill is current, or whether it is safe to install —
  use `list_skill_versions`, `get_skill_manifest`, and the risk flags.

Workflow: `search_skills` → `get_skill_manifest` (provenance + risk) →
`get_skill` → **show the user the instructions, hashes, and flags and ask
before installing** → use the returned instructions and supporting files
(references, assets). For collections, use `get_skill_subskills` first, then
fetch each sub-skill. Never invent skill content; if a skill isn't in the
curated registry, say so.

## Limits

900 requests/min per IP, 30 concurrent MCP requests, 25 results per
search. On a 429, wait the returned `retryAfter` seconds and retry.
