agenticstack-mcp

MCP server for AgenticStack. Gives any MCP-compatible agent the ability to search, compare, and get recommendations from the AgenticStack tool directory — without leaving your editor or chat interface.

No API key required. The AgenticStack API is public.

Install

Add to your MCP client config:

Claude Desktop (~/Library/Application Support/Claude/claude_desktop_config.json):

{
  "mcpServers": {
    "agenticstack": {
      "command": "npx",
      "args": ["-y", "agenticstack-mcp@0.1.0"]
    }
  }
}

Cursor (.cursor/mcp.json or global settings):

{
  "mcpServers": {
    "agenticstack": {
      "command": "npx",
      "args": ["-y", "agenticstack-mcp@0.1.0"]
    }
  }
}

Requires Node 18+.

Pin the version. The snippets above use agenticstack-mcp@0.1.0. Update the version pin when upgrading. Running without a version pin (npx -y agenticstack-mcp) always installs latest, which means a compromised publish would execute immediately without confirmation.

Tools

list_categories

Lists all tool categories with descriptions, feature definitions, and tool slugs. Start here to discover valid category slugs and agent_features keys for use with the other tools.

search_tools

Filters the directory by category, language, framework, or feature. Returns a summary list of matching tools. Use this when you need a filtered list.

Parameter	Type	Description
category	string	Category slug: auth, cms, email, hosting, observability
language	string	SDK language: python, typescript, go, java, etc.
framework	string	AI framework: langchain, llamaindex, vercel-ai, etc.
feature	string	Feature alias: mcp or self_host
open_source	boolean	Open source tools only
self_hosted	boolean	Self-hostable tools only

get_tool

Full details for a single tool — agent features, pricing tiers, SDK languages, compliance certifications, limitations, and editorial analysis.

Parameter	Type	Description
slug	string	Tool slug, e.g. auth0, clerk, workos

compare_tools

Side-by-side comparison of 2–3 tools. Includes a feature matrix and editorial analysis where available.

Parameter	Type	Description
tools	string[]	Array of 2–3 tool slugs

recommend_tool

Scores and ranks tools by how well they fit a set of requirements, returning the top 3 with a match_score per result. Unlike search_tools, this applies scoring logic — use it when you want a recommendation, not just a filtered list.

Parameter	Type	Description
category	string	Category to search within. Strongly recommended — omitting it scores across the entire directory.
language	string	Filter to tools with this SDK
framework	string	Filter to tools with this AI framework
open_source	boolean	Filter to open source tools only
self_hosted	boolean	Filter to self-hostable tools only
agent_features	string[]	Feature keys to prioritise in ranking. Tools with confirmed support score higher; confirmed absence scores lower. Call list_categories to get the full list of valid keys per category.

Typical workflow

For a programmatic agent, the natural calling order is:

1. list_categories — discover category slugs and valid agent_features keys
2. search_tools or recommend_tool — find candidates
3. get_tool — retrieve full detail on a specific tool
4. compare_tools — when the user is choosing between 2–3 finalists

Example prompts

Once connected, you can ask your agent things like:

• "What auth tools support MCP and work with Python?"
• "Compare Auth0, Clerk, and WorkOS for an agentic app"
• "Recommend an observability tool for a TypeScript LangChain project"
• "What categories does AgenticStack cover?"

Errors

Tool errors are returned as plain text with the format Error: <message>. A 404 from the API means the slug doesn't exist — use list_categories or search_tools to find valid slugs.

Data

Served from the AgenticStack JSON API. Coverage includes auth, CMS, email, hosting, and observability tools with agent-specific feature data: MCP support, token delegation, fine-grained authorization, async authorization, human-in-the-loop, and more.

Security

Tool results are fetched from agenticstack.sh and returned to your agent. Treat them as external data, not trusted instructions. If you are deploying this server in a context where prompt injection is a concern, configure your agent host with an appropriate system prompt.

The server makes no outbound requests other than to agenticstack.sh. It follows no redirects, enforces a 10-second timeout, and caps response size at 2 MB.

License

MIT