Project-local MCP server for cross-provider agent discovery, current-work status, and durable direct messages.
Let agents DM and coordinate each other directly.
Agent DMs gives the agents you already run a shared, persistent inbox.
https://github.com/user-attachments/assets/e4e7955b-4e56-47c7-be71-a7066bc94a9a
Review draft: the supplied film contains illustrative usage and popularity figures, not verified product results.
Stop copying messages from one agent to another. Connected agents can find peers, share what they are working on, send a DM, and pick up a reply—all through one project-local MCP server.
Use the agents and models you already have. Agent DMs provides the messaging, not the model execution: it does not launch agents or manage their processes. HTTP clients and the stdio bridge connect to the same running daemon and mailbox.
Reading a DM does not delete it. The receiving agent claims a leased batch, handles the messages, then explicitly acknowledges the exact messages it handled. If a claim expires before acknowledgement, the message can be claimed again. Replies are messages too—not automatic proof that the original task is complete.
Use this for a question, a handoff, or a review between existing agents. The protocol also supports explicit Manager/Worker questions, completion reports and review replies within their existing permissions.
Experimental 0.1 series. Check Releases for published versions; until one exists, use an authorized source checkout. Linux and Python 3.11/3.12 are tested. MCP HTTP/stdio interoperability is tested with the official SDK; native Claude Code and Codex sessions still need the client acceptance check. Provider labels in the demo d
From the project README.
Add the radar badge to your README — it shows your project was picked up by MCP Radar and links to this page:
[](https://mcp.liqiwa.com/s/49-Agents--agent-dms.html)
The top new MCP servers of the week, every Monday. No spam, unsubscribe anytime.