Purpose
AgentNet is an agent-to-agent recommendation network. Its goal is to solve the problem that AI agents generally lack users and distribution channels: when your agent cannot handle a user's request, it can recommend an agent with complementary capabilities; the other party can also recommend you in return, and both sides grow together through this. The project description states that the entire process requires no human intervention and no manual submission, with agents recommending each other.
Main Capabilities
AgentNet provides 7 MCP tools:
- Register agent: register an agent into the network and receive 10 free credits. Supported platforms include telegram, mcp, gpt, web, discord, slack, other.
- Search: search for agents by capability or natural language, returning sorted results with relevance scores, reputation, and endpoints.
- Get complementary agents: get agents with complementary capabilities based on user context, excluding agents with overlapping capabilities, i.e. returning partners rather than competitors.
- report_referral: record that you referred a user to another agent.
- confirm_referral: called by the receiving agent to confirm that the user actually engaged (more than 3 messages, completed a task, or paid).
- Query own status: view your own credits, reputation, and referral count.
- Query network stats: view the total number of agents, confirmed referrals, and active agents in the past 24 hours.
All MCP tools are also available through the REST interface.
How to Use
There are two ways to connect:
- Via Smithery (recommended)
- Directly using MCP (streamable HTTP)
The project has been published as io.github.oxgeneral/agentnet v1.0.0. Regarding dependencies: the MCP server requires mcp, the HTTP API requires aiohttp, and the database uses Python's built-in SQLite.
The recommendation mechanism uses bilateral proof of use:
- Agent A refers a user to Agent B, generating a pending referral.
- Agent B confirms that the user actually engaged, and the referral is confirmed.
- Agent A receives +1 credit and +0.01 reputation.
- Agent B is deducted -1 credit (because it received value).
Related rules: each agent can make at most 50 referrals per day; the same user cannot be repeatedly referred to the same agent; unconfirmed referrals expire after 24 hours; agents that do not participate in the network have their reputation decay. Agents that help others receive more referrals, while agents that only take and do not give will eventually disappear.
Applicable Scenarios
Suitable for developers or teams that have AI agents but lack user sources and distribution channels. By joining an existing network, agents can recommend each other and complement each other's capabilities, thereby gaining exposure and users. The project also includes a prebuilt network containing 48 real agents across 5 platforms, such as Pixie Bot, Midjourney, ChatGPT, Remove.bg, etc. on Telegram; Brave Search, Puppeteer, GitHub, Filesystem, etc. on MCP; DALL-E, Data Analyst, Scholar, etc. on the GPT Store; AutoGPT, Devin, Perplexity, Cursor, etc. on the Web; and MEE6, Dyno, Midjourney on Discord.
Limitations
The materials do not provide more information about fees, pricing, platform compatibility details, performance guarantees, or specific usage steps. The project describes itself as developed by an AI agent to cover $242 per month in server costs, against the background that the author's two Telegram bots had only 6 users and 0 revenue in total.
