Skip to content

Agent Support

Memet Zx edited this page Nov 15, 2026 · 2 revisions

Agent Support

Mools is built to be used by AI agents as well as people. The same tools are available four ways.

Surface What it is Where
llms.txt A short summary of the site with links to every tool https://maling-it.github.io/tools/llms.txt
tools.json All pages plus the JSON Schema of every callable tool https://maling-it.github.io/tools/tools.json
WebMCP Tools registered in the browser through navigator.modelContext Every page, in browsers that support it
MCP server A stdio Model Context Protocol server mcp/server.js in the repository

Pages also use labeled form fields and SoftwareApplication JSON-LD, so browser agents can drive the UI directly.

Running the MCP server

git clone https://github.com/maling-it/tools.git
cd tools
bun install
bun run mcp        # or: node mcp/server.js

Add it to your MCP client, for example Claude Desktop or Cursor:

{ "mcpServers": { "mools": { "command": "node", "args": ["/absolute/path/to/tools/mcp/server.js"] } } }

The site has an MCP Config Generator tool that writes this snippet for Claude Desktop, Cursor, VS Code and Claude Code.

What the server offers

  • Tools (79). Encoding, converters, hashes, generators, calculators, formatters, bulk helpers and AI helpers. All are marked read-only. A few HTML and XML tools use jsdom, an optional dependency.
  • Prompts (24). The prompt library as MCP prompts. Every {{placeholder}} becomes an optional argument.
  • Resources. mools://catalog lists the web tools and mools://prompts contains the prompt library.

Using the prompts in Claude Code

Claude Code turns each MCP prompt into a slash command named /mcp__mools__<prompt>:

/mcp__mools__fix-bug                 Claude asks for the symptom, what you expected and what happened
/mcp__mools__security-audit login-endpoint
/mcp__mools__commit-message

Claude Code passes one word per argument, so give a single word (or join words with hyphens) or leave the arguments out. A placeholder that stays open is listed at the end of the prompt ("Still open: ..."), and the model takes it from the conversation or asks you. For full sentences, ask Claude to call the search_prompts tool with a values object.

What results look like

  • Errors say what to fix, for example The input is not valid JSON: ... or "kg" is not a length unit. Use one of: m, km, ....
  • A search without matches says so (No email addresses found.). Tools that return JSON lists return [].
  • No tool answers with an empty response: the fallback is (The result is empty.).
  • Arguments are checked against each tool's schema, and a call that runs for more than 15 seconds is stopped.

Good to know

  • Everything runs locally. The server makes no network calls.
  • Tool results are plain text and usually ready to paste.
  • Errors come back as MCP error results, not crashes.
  • Tools such as the password generator return different output on every call, and their annotations say so.

Adding a tool for agents

Add it to src/lib/agentTools.js. See Adding a Tool.

Clone this wiki locally