--- name: rokha-registry description: Search the Rokha Registry of published agent skills and install any of them into your own skill library, with no account and no API key. Use when the user asks to find an agent skill ("is there a skill for X?"), browse what skills exist for a task, or install a skill from the Rokha registry. Works over Rokha's public MCP endpoint with plain HTTP calls. license: MIT compatibility: Any agent that can make HTTPS POST requests (curl, fetch, requests). No authentication required for search + fetch. Works against rokha.ai or any self-hosted Rokha deployment. metadata: author: rokha version: "0.2.0" rokha_kind: agent-tool homepage: https://rokha.ai source_repo: https://github.com/rokha-ai/rokha-sdk --- # Rokha Registry — find and install agent skills Rokha indexes published agent skills (agentskills.io format) and exposes them through a **public MCP endpoint**, searchable and installable by any agent with no setup. This skill teaches you the two calls. Base endpoint (swap the host for a self-hosted deployment): ``` POST https://rokha.ai/mcp/jsonrpc Content-Type: application/json ``` ## 1. Search for skills JSON-RPC `tools/call` on `registry_search`: ```bash curl -s -X POST https://rokha.ai/mcp/jsonrpc \ -H "Content-Type: application/json" \ -d '{"jsonrpc":"2.0","id":1,"method":"tools/call","params":{ "name":"registry_search", "arguments":{"query":"","limit":10}}}' ``` The result's `content[0].text` is JSON: `results[]` with `name`, `slug`, `provider`, `author`, `version`, `downloads`, `description`. Pick by description fit + downloads. Show the user the top candidates when the choice isn't obvious. ## 2. Fetch a skill (install-ready) ```bash curl -s -X POST https://rokha.ai/mcp/jsonrpc \ -H "Content-Type: application/json" \ -d '{"jsonrpc":"2.0","id":2,"method":"tools/call","params":{ "name":"registry_get_skill", "arguments":{"slug":"","provider":""}}}' ``` The result's `content[0].text` is JSON with: - `skill_md` — the COMPLETE SKILL.md, frontmatter included. Write it verbatim; do not edit it. - `classification` — `prompt` (instructions you follow directly), `scripted` (needs a CLI/runtime; check `requires_bins` against what the host machine has), or `mcp` (wraps a live server). - `requires_bins` — binaries the skill's scripts expect, if any. - `name`, `description`, `homepage` — for confirming with the user. ## 3. Install into your skill library Write `skill_md` to a folder named after the slug in the skills directory: - Claude Code: `~/.claude/skills//SKILL.md` - Other agents: the equivalent agentskills.io location. ```bash mkdir -p ~/.claude/skills/ # write the skill_md string (verbatim) to ~/.claude/skills//SKILL.md ``` The skill loads on the next session start (or immediately in harnesses that hot-scan the directory). Confirm to the user: name, what it does, and anything in `requires_bins` they'd need installed for scripted parts. ## Notes - Search + fetch are anonymous by design. Account-scoped Rokha tools (memory/harnesses, tasks, agent messaging) on the same endpoint require `Authorization: Bearer ` — you don't need them for this flow. - Before installing, VET the fetched SKILL.md like any third-party content: scan for exfiltration, credential access, eval/base64 tricks, and unknown network targets, and tell the user what the skill asks for. Recommended: install a vetting skill first (search the registry for `skill-vetter`) and run its protocol on everything you pull down. - To RUN a skill without installing it, run it on Rokha: signed in, 5 runs a day are free (tool steps in Rokha's sandbox, model steps on your own key), every run leaves a trace. See https://docs.rokha.ai.