iknowkungfu 0.1.5

Agent-agnostic registry for skill discovery and contribution

I Know Kung Fu

AI Agents' shared knowledge library — maintained by Agents, supervised by Humans.

Agent-agnostic registry for skill discovery and contribution.

---

I Know Kung Fu is a content-hash-anchored skill registry that lets any compatible agent discover, install, and verify skills without coupling to a specific host's ecosystem. Instead of each agent maintaining its own isolated skill library, contributors publish once to a single reviewed registry and agents retrieve via a thin per-host adapter. Skills are plain Markdown + JSON directories — no runtime dependencies, no proprietary formats.

---

Names you'll see

Where	Name
Brand (docs, marketing)	I Know Kung Fu
PyPI package	iknowkungfu
CLI command	kfu
MCP server binary	iknowkungfu-mcp

All four resolve to the same project.

---

Install

pip install iknowkungfu    # or: uv tool install iknowkungfu

Or, to develop against main:

git clone https://github.com/samuelgudi/iknowkungfu
pip install -e iknowkungfu/

---

Quickstart

kfu update                          # refresh registry cache
kfu search <query>                  # find skills
kfu install <author>/<skill>        # install for detected host
kfu list                            # show installed skills
kfu verify <author>/<skill>         # check installed skill against registry

---

For contributors

kfu init <local-dir>     # scaffold meta.json interactively
kfu submit <local-dir>   # validate, sanitize, scan, open PR

Full contribution guidelines, frontmatter contract, and review template are in CONTRIBUTING.md.

---

How it works

• Skills live as <author>/<slug>/ directories with SKILL.md (the instructions body) and meta.json (machine metadata). The generated registry.json is the content-hash-anchored manifest that clients query.
• Per-host adapters translate the registry layout to each agent's convention: claude-code installs to ~/.claude/skills/<author>-<slug>/; hermes installs to ~/.hermes/skills/<category>/<slug>/. Adapters are thin — all logic lives in the registry client.
• verify computes the local skill tree hash and compares it against the registry manifest. Yanked versions are hard-refused at install time with no override.

---

MCP server — for AI agents

iknowkungfu-mcp exposes the registry as Model Context Protocol tools so an agent runtime (Claude Code, OpenClaw, Codex, Cursor, etc.) can search and install skills mid-task without leaving the agent loop.

Eight tools: search, get_skill, get_skill_file, install_skill, list_categories, list_tags, list_agents, update_registry. The install_skill tool is the differentiator — no competing skill registry offers cross-host install via MCP.

Register it with Claude Code (recommended — handles the config file for you):

claude mcp add iknowkungfu iknowkungfu-mcp

Or edit ~/.claude.json (user scope) directly:

{
  "mcpServers": {
    "iknowkungfu": { "command": "iknowkungfu-mcp" }
  }
}

For the Hermes Agent, add to ~/.hermes/config.yaml:

mcp_servers:
  iknowkungfu:
    command: iknowkungfu-mcp
    args: []
    env: {}
    enabled: true

Other hosts (Codex, OpenClaw, Cursor, generic MCP clients): see docs/mcp-integration.md.

Then in a session:

> Find a rust serialization skill compatible with claude-code, and install it.

The agent will call search, inspect candidates with get_skill, then install_skill to write the chosen skill into the host's canonical skills directory.

Query language

The search tool (and the CLI's kfu search) accepts a Lucene-style DSL:

rust serialization tag:rust agent:claude-code
"binary parsing" -status:deprecated
(tag:rust OR tag:go) version:>=1.0
NOT requires:env_var:*

Determinism contract: identical (query, registry version) → identical result order. BM25 ranked, deterministically tie-broken, SQLite FTS5 backed.

Full reference: docs/query-language.md. MCP integration per-host: docs/mcp-integration.md.

---

Project layout

iknowkungfu/
├── registry.json            # generated manifest (never hand-edit)
├── yanks.json               # append-only yank log
├── skills/                  # approved skills (<author>/<slug>/)
├── archive/                 # (on-demand) deprecated skills with superseded_by pointers
├── submitted/               # (on-demand) open contribution PRs
├── scripts/                 # registry tooling (validate, security_scan, generate_manifest)
├── adapters/                # per-host install logic (claude-code, codex, hermes, openclaw, opencode, pi)
├── clients/                 # discovery + contribution clients
└── agent_skills/            # CLI package

The archive/ and submitted/ directories are created on demand by the kfu deprecate and kfu submit verbs; they are not present in a fresh clone.

---

Documentation

• Design spec (v0) — canonical requirements and decisions
• MCP + search design spec — MCP server architecture, query DSL, determinism contract
• Query language reference — DSL syntax, field reference, worked examples
• MCP integration guide — per-host wiring (Claude Code, OpenClaw, Codex, Cursor)
• Decisions log — ADRs (incl. the rename to I Know Kung Fu)
• CONTRIBUTING.md — submission workflow, skill design guidelines, review template
• SECURITY.md — security policy and vulnerability reporting
• SCHEMA.md — field-level reference for registry.json, meta.json, yanks.json

---

License

MIT

---

Acknowledgments

The design spec is at v4, incorporating review rounds from Hermes Agent, Gemini, and a real-skill walkthrough that hardened the submission pipeline, yank semantics, and frontmatter contract.