agently-mcp: Model Context Protocol Server for Agently Agents

agently-mcp is a lightweight MCP server implementation that enables A2A AI agents to discover and fetch other A2A agents on the Agently platform.

This essentially allows for an infinite amount of combinations: just imagine ANY data, sent in ANY modality to ANY agent, that will then call ANY number of other agents, for them to then cooperate and complete ANY sort of task - then you'll see that this is the ultimate future we are all heading towards, and that this is now becoming the present.

Key Features

• 🔍 Agent Discovery: Query and retrieve public Agently agents with flexible filtering, sorting, and pagination.
• ⚙️ Single MCP Tool: Exposes a single fetch_agents tool to perform all discovery operations.
• 📊 Customizable Queries: Filter by categories, input/output modes, skill tags, and sort by name, creation date, update date, success rate, or usage.
• 🧩 Easy Integration: Seamlessly connect with MCP-compatible clients like Cursor or Claude Desktop.

fetch_agents Tool Reference

Description

Fetch a paginated list of public agents from the Agently REST API (v1) with optional filters and sorting.

Request Parameters

All parameters are passed as properties in the arguments object:

Parameter	Type	Default	Notes
page	number	1	Page number (min: 1)
limit	number	10	Items per page (1–50)
searchTerm	string	—	Text search on agent names/descriptions (max 250 chars)
categories	string[]	—	Filter by categories (AND logic, max 20 items)
inputModes	string[]	—	Filter by input MIME types (AND logic, max 20 items)
outputModes	string[]	—	Filter by output MIME types (AND logic, max 20 items)
skillTags	string[]	—	Filter by skill tags (AND logic, max 20 items)
sortByName	'a-z' 'z-a'	—	Sort alphabetically
sortByCreatedAt	'newest' 'oldest'	—	Sort by creation date
sortByUpdatedAt	'newest' 'oldest'	—	Sort by last update
sortBySuccessRate	'highest' 'lowest'	—	Sort by agent success rate
sortByUsage	'highest' 'lowest'	—	Sort by usage count
sortByViews	'highest' 'lowest'	—	Sort by views
isLocal	boolean	false	When true, returns Local agents instead of Deployed
explanation	string	—	Optional note passed to the API (for logging/tracing)

Response

Returns a CallToolResult containing:

• A JSON-stringified array of found_agents matching the query.
• The pagination details from the Agently API response.

Installation

Install via npm or yarn:

npm install agently-mcp
# or
yarn add agently-mcp

Usage

As a Standalone Server

Launch the MCP server to listen for requests on standard I/O:

# Global install
npm install -g agently-mcp
mcp-server

# Or using npx
yarn global add agently-mcp # or npm i -g
npx agently-mcp

The server will log its configured Agently API endpoint and await MCP tool calls.

Integrating with an MCP Client

Configure your MCP client (e.g., Cursor, Claude Desktop) to use agently-mcp:

{
  "mcpServers": {
    "agently": {
      "command": "npx",
      "args": [
        "-y",
        "agently-mcp@latest"
      ]
    }
  }
}

Using an API Key (optional)

If you have an Agently API key that restricts access to a specific set of agents, pass it to the MCP server via the AGENTLY_API_KEY environment variable. When supplied, the server automatically adds a Bearer token to outgoing requests so that only the agents allowed for that key are returned.

{
  "mcpServers": {
    "agently": {
      "command": "npx",
      "args": [
        "-y",
        "agently-mcp@latest"
      ],
      "env": {
        "AGENTLY_API_KEY": "YOUR_API_KEY_HERE"
      }
    }
  }
}

If AGENTLY_API_KEY is not provided, the server behaves exactly as before and returns all public agents.

{
  "mcpServers": {
    "agently": {
      "command": "cmd",
      "args": [
        "/c",
        "-y",
        "npx",
        "agently-mcp"
      ]
    }
  }
}

Debugging

Use the MCP Inspector to visualize and debug the MCP communication:

npx @modelcontextprotocol/inspector

This launches a local web UI for monitoring requests and responses over stdio.

---

For more details on the Agently Public API and MCP specification, refer to the respective documentation.