Devotd Github
NewAI agents

chekusu/yobiru v0.7.15 — Tailscale-based agent communication layer

chekusu/yobiru is a Python communication layer that connects personal agents across multiple computers over Tailscale so they can exchange messages and files.

Repositorychekusu/yobiru ↗

chekusu/yobiru is a Python communication layer that connects personal agents across multiple computers over Tailscale so they can exchange messages and files.

GitHub description: Let your agents talk: messages, data, files and wake-ups over Tailscale.

stars
★ 381
Forks
53
Language
Python
License
MIT
Latest release
v0.7.15
Read original source ↗View on GitHub ↗

chekusu/yobiru is a communication layer that lets personal agents on multiple computers and in cloud workspaces exchange messages, structured data, and files over Tailscale, and send wake-up signals. The repository's latest release is v0.7.15, and the provided information does not specify what exactly changed in that release. According to the README, agents can respond to each other or request a wake-up through this layer while keeping their existing interfaces. It is written in Python and released under the MIT license.

The role of this project is to layer agent-to-agent communication on top of Tailscale without a separate service. The transmission method described in the README is the tailmsg protocol v1, and the implementation uses only the Python standard library. Files are staged on the sender side behind links that expire after 24 hours by default, and the receiver fetches the file and verifies its size and SHA-256. Yobiru does not operate a service that stores messages or attached files. Connection peers are identified with tailscale whois and filtered out by local access policies. So file delivery and peer verification are in scope of the communication layer, not just message delivery.

That said, the evidence that this design has been sufficiently validated in real environments is still limited. The README states that it does not provide an MCP server and instead relies on integration with host skills, and that end-to-end integration has not been validated either. No performance benchmarks are presented either. So what can be confirmed at present is an implementation and an approach for setting up agent communication over a private network, and its maturity for operational use still requires further validation.