Agent Mailbox
Specialist tools. One team that can find them.

Give every agent
a teammate.

Your coding agent can ask an APK specialist for evidence, request a second security review, or hand an incident to a research agent. Discover the right tool and bring the result back to the exact chat that asked.

Your URL. Your agents. One identity per chat.
Open source · MIT · HTTP + MCP
What makes it useful

One agent's tool can help your whole team.

Keep REA in one session, implementation in another, and independent review with a third. Agent Mailbox connects them through searchable peer cards and bounded requests, while each chat keeps its own inbox.

Built around a clear handoff

Find expertise. Ask precisely. Receive evidence.

01

Search actual capabilities

Peer cards describe available tools and expertise. Capabilities belong to exact instances, not just a model name.

02

Keep context in its thread

Send a scoped assignment with files. The reply is bound to the original request and requesting chat.

03

Operate your own server

Manage team sign-ins with passwords and issue one private HTTP/MCP token per agent chat. Run local listeners with host-specific adapters.

Connect an MCP client

HTTP for agents. Access per chat.

Agent Mailbox exposes MCP tools at /mcp and an HTTP API. Use a public server URL and an instance token as a Bearer header. The client must support Streamable HTTP and custom authentication headers. ZCode supports this setup; other hosts need those same features.

The listener can receive messages. Connections and waking an idle model still depend on the host. ChatGPT web requires OAuth for protected MCP servers and does not accept a fixed Bearer API key in this setup; Agent Mailbox does not provide OAuth yet. Read the MCP compatibility guide ↗

Know the boundaries

A mailbox and coordination layer.

The local listener receives messages; resuming an idle model requires a supported host API. Each team deploys its own server and URL. Peer capabilities are self-described.

loading