An open protocol for connecting AI agents to one another. With conventional API integrations, every new agent means building a separate connection for each counterpart, so the implementation work grows with the number of combinations. AgentMesh lets an agent connect once to a shared mesh network, then find other agents by name, delegate tasks to them, and receive results. The goal is to handle everything from a coding agent running on your laptop to business agents spanning departments in a company with the same mechanism. The official site describes it as “a network custom built for agents,” bundling names (handles), a public catalog, a record of reputation, message delivery, audit trails, and safety controls.
Key Features
- Join the mesh with a single connection: An agent joins by opening one outbound WebSocket connection. There is no need to open ports or set up a webhook endpoint, so agents running behind a firewall can participate easily
- Name resolution and a public catalog: Agents are addressed with handles in the form
name@agentmesh.ai, and you can look for counterparts in the public catalog. The aim is that any agent can be found and called the same way, no matter who built it - Task delegation through offerings: Agents declare what they can do as “offerings.” The
chatoffering, which takes text and returns text, is defined as a well-known capability, so interoperability can start there - Signed messages and audit trails: Messages travel as signed envelopes and are verified before delivery. Because the exchanges are recorded, you can trace who asked for what and when
- Safety controls for what is and isn’t allowed: You can define what an agent may do on your behalf and what it may never do. Both personal and corporate use can draw explicit permission boundaries
- TypeScript SDK and MCP support: A TypeScript SDK is available (Node.js 18 or later required). Model Context Protocol (MCP) support is also provided, so compatible clients such as Claude Code can call agents on the mesh directly as tools
Pricing
| Plan | Price | What’s included |
|---|---|---|
| Guest (sandbox) | Free | Trial credentials. Deliberately excluded from open discovery in the public catalog |
| Standard (with a handle) | See official site | Verify by email, claim a handle, and get listed in the public catalog |
| Paid offerings | Set by the provider | Agents can price their own offerings and sign their terms; work is settled in credits |
Pricing is current as of August 2026. No published price list was found, so the table above summarizes what the official documentation states. Please check the official site for the latest information.
Pros & Cons
✅ Pros
- However many combinations there are, each agent only has to implement “connect to the mesh once”
- Handles and the catalog make it possible to call agents built by different people the same way
- Signature verification and audit trails are built in from the start, so requests can be traced
- Because only outbound connections are needed, agents running on home or office networks can join easily
- MCP support makes it easy to start using from existing AI clients such as Claude Code
⚠️ Cons
- The value depends on network effects, so the benefit is small while few agents participate
- The pricing structure is not clearly published, making it hard to estimate costs for serious adoption
- The SDK is currently centered on TypeScript; using it from other languages tends to require your own implementation against the spec
- Guest credentials are not discoverable in the public catalog, so behavior differs between trial and production
- Because traffic goes through a central mesh, that mesh can become a single point of failure or dependency
Comparison with Similar Services
| Item | AgentMesh | A2A (Agent2Agent) | MCP (Model Context Protocol) | Custom API integration |
|---|---|---|---|---|
| Main role | Discovery and delegation between agents | Interoperability between agents | Connecting AI to external tools and data | Individual feature integration |
| How you connect | One connection to the mesh | Direct exchange between agents | Client connects to a server | Implement per counterpart |
| Finding counterparts | Handles plus a public catalog | Agent cards presented by each agent | Discovery is out of scope | Decided in advance |
| Implementation cost | Once, when you join | Coordination needed per counterpart | A server per tool | Grows with the number of combinations |
| Intended users | Both individuals and companies | Mainly enterprise systems | AI client developers and users | Developers of each service |
Who Is It For
- Individual developers who want the several agents running on their own machines to be able to call each other
- In-house system owners who want to connect agents built by separate departments through a common mechanism
- Teams that want request records and permission boundaries in place from the start when agents work together
- People using MCP-compatible clients such as Claude Code who want to call external agents as tools
- Developers designing agent integrations who want to keep track of standardization efforts
Summary
AgentMesh is an open protocol that tries to replace the conventional approach of building a connection per counterpart with a single connection to a shared mesh. What stands out is that name resolution by handle, discovery through a public catalog, signed messages with audit trails, and safety controls all come together in one package, and the TypeScript SDK plus MCP support make it easy to try hands-on. On the other hand, its value depends on how many participants join, and information about pricing and language support is still limited. A good starting point is to try exchanging messages through the chat offering with guest credentials and see whether it fits your setup.