agentmods.dev

Use the catalogue from your agent

A person browses this site. An agent should not have to: tell it once how to search 649,047 measured mods and it can find, compare and install them itself — choosing by token cost, scan grade and originality instead of by the first search hit. There are two ways in; pick by whether your agent has a shell.

Agent with a shell → the skill

Claude Code, Codex, and anything that can run npx. The skill teaches the flow: search the API, compare the measurements out loud, install with the CLI. One file, 60 tokens per session.

Set up the find-mods skill ↓

Agent without a shell → the MCP server

Claude Desktop and any MCP client. Two tools over stdio: search_mods and mod_info, both returning the measurements with every result.

Set up the agentmods MCP server ↓

The find-mods skill

A single SKILL.md that gives the agent a procedure, not just an endpoint: search /api/search (names and plain intent both work — “playwright” and “make my agent cheaper” alike), then choose by the measurements: originals over copies, scan grade A/B over the rest, fewer tokens per session when candidates tie, and say the numbers to the user while doing it. Install is npx agentmods add, which itself prints the cost and the scan before writing anything.

Install the skill into a project:

npx agentmods add skills/iosdevsk/find-mods/find-mods

Claude Code gets it in .claude/skills/, Codex in .agents/skills/. Then just ask naturally:

  • “find me something for code review and install it”
  • “I want a hook that runs tests before every commit”
  • “my sessions feel expensive — find out why and fix it”

The skill also covers updating (npx agentmods update --check exits 1 in CI when an installed mod changed upstream) and refuses to install a plugin behind the user’s back — plugins go through Claude Code’s own /plugin flow, and the skill says so instead of pretending.

Source: github.com/iOSDevSK/find-mods (MIT)

The agentmods MCP server

For agents that cannot run commands, the catalogue speaks Model Context Protocol. Zero dependencies, stdio transport, Node ≥ 18. Add to .mcp.json in a project, or to Claude Desktop’s MCP configuration: 

{
  "mcpServers": {
    "agentmods": { "command": "npx", "args": ["-y", "agentmods-mcp"] }
  }
}

Two tools:

  • search_mods { query, type?, limit? } — hybrid search over the whole catalogue; every result carries tokens_per_session, security_grade, origin (original / copy), repository stars, the page URL and the ready install command. Duplicate clusters collapse to one result — a skill forked into two hundred repositories is one answer.
  • mod_info { slug } — one mod in full: description, licence, token cost per session and once invoked, scan grade with findings count, the files it ships, and per-agent install commands for Claude Code, Codex, Gemini CLI, Cursor and GitHub Copilot.

The same server also ships inside the CLI — npx agentmods mcp — so a machine that already uses the agentmods package needs nothing extra.

Source: github.com/iOSDevSK/agentmods-mcp (MIT) · npm: agentmods-mcp

Why the measurements travel with every answer

Any catalogue can return search hits. The point of this one is that the model doing the choosing sees, for every candidate, what it will cost in tokens on every future session, what a security scan found and where, and whether it is the original or the two-hundredth copy. An agent given those numbers picks the 60-token grade-A original over a 2,000-token unscanned clone — and can tell its user why. That is the whole difference between installing something and installing the right thing.

Plain HTTP is always there too: GET /api/search?q=…&type=…&limit=… returns the same measured results as JSON, no key required.