The ToolRouter skill is a portable Markdown workflow for using the catalog through MCP or an agent runtime. It tells the assistant to discover an operation, inspect the schema and cost, execute it, resolve queued jobs and return the actual result. It does not authenticate an account or authorise spending.
Install the skill
From the project where the agent works:
pnpm dlx skills add Humanleap/agent-skills --skill toolrouterSelect your supported client when the installer asks. The public skill repository contains separate product folders; choose toolrouter rather than treating the entire collection as a single skill.
OpenClaw also documents a native, source-qualified installer:
openclaw skills install skills-sh:Humanleap/agent-skills/toolrouterWhere does SKILL.md go?
| Client | Documented skill location or import |
|---|---|
| Claude Code | .claude/skills/toolrouter/SKILL.md or the user skills folder |
| Codex | .agents/skills/toolrouter/SKILL.md |
| Cursor | .agents/skills or .cursor/skills |
| Gemini CLI | .gemini/skills or .agents/skills |
| Kimi Code | .kimi-code/skills or .agents/skills |
| Grok Build | .grok/skills or its user skills folder |
| Hermes | ~/.hermes/skills/toolrouter/SKILL.md |
| OpenClaw | Workspace or user skills folder; native installer above |
| Claude and Cowork | Account skill upload through Customize > Skills |
| Paperclip | Company skill import with the appropriate company permission |
These locations describe the named runtime. A local folder does not automatically install a cloud plugin or sync into another agent account. ChatGPT and Perplexity Computer use the remote connector guide here; a generic local SKILL.md import is not confirmed for those surfaces.
For Claude account upload, ZIP the toolrouter folder with SKILL.md at its root and its supporting resources, then use Upload a skill when available. The account needs the features and policy required to run that skill.
Connect the tools separately
Add https://api.toolrouter.com/mcp using your client guide, then complete the advertised sign-in or account-claim flow. Claude chat and Cowork guides use the Anthropic-specific profile at /mcp/anthropic, whose catalog excludes standalone AI image, video and synthetic audio generation.
For a shell runtime, follow the CLI guide and configure the intended hosted account. Keep provider credentials and API keys in the runtime secret configuration, outside the skill file and chat transcript.
What the workflow teaches
- Discover the requested outcome rather than inventing an operation.
- Read the returned tool, skill, schema, price and account requirements.
- Check the budget and upstream account for paid or account actions.
- Execute the discovered operation and resolve a returned asynchronous job.
- Return source links, completed asset URLs or the provider action result, with errors and dates visible.
The provider reference describes public integrations, including Firecrawl, Exa, DataForSEO, fal.ai, ElevenLabs, Companies House, SEC EDGAR and the World Bank. A provider page is context; the live catalog and returned schema determine what your account can call.
Test the installed workflow
Read the ToolRouter skill. Discover a free operation for current weather in London, run it and show every returned field. Do not buy credits or change any connected account.
An installed skill label is not a completed provider call. Check that the agent actually uses discovery, chooses a real operation and returns its output. For paid work, review the discovered price before running a batch.
Updating and troubleshooting
Review SKILL.md and its resources before refreshing an installed copy. Restart or reload the client after changing skill files. If the skill is visible but tools are missing, inspect the MCP connection; if tools work but a provider action fails, inspect account requirements rather than reinstalling instructions.
Sources: Claude Code skills, Codex skills, OpenClaw skills.