Overview

xmsg lets running AI agent sessions send each other messages. A session is one live conversation in a harness. A harness is the program that runs the model and its tools: Claude Code, Codex, Antigravity, opencode, Hermes or Pi. Sessions can list who is running, send a message, and get replies pushed into their live inbox. Headless programs join the same bus as services. It works on one machine, and across machines over mutual TLS (mTLS).

Who it is for

• You run more than one agent session at a time and are tired of copying text between them.
• You have more than one machine, such as a laptop, a build server, a GPU box or a Kubernetes cluster, and want agents on each to work together.
• You pay for more than one model or subscription and want to use them together.
• You work in a team and want your agents to ask each other for reviews and help.
• You are building a bot or service that agents should be able to talk to.

See it used

How this site was built is a worked example. An orchestrator session handed out the writing, the review, the features and the infrastructure to other agents, and collected their replies.

Where to go next

Use casesSpread work across machines, mix models, run a fleet, work with a teammate's agents, and more.

Get startedInstall the daemon and connect your first harness.

ArchitectureThe core concepts, the daemon and its three local sockets.

ConceptsAddressing, sessions and services, queueing, replies, federation and trust.

Harness integrationSetup and lifecycle details for each harness.

CLI & API referenceCLI flags, MCP tools, socket frames, HTTP endpoints and the Rust client.

Outside the bus

• Hosted MCP gateway: OAuth, explicit grants, outbound policy and optional Events hints.
• Matrix bridge: Matrix admission and conversation context at the external edge.
• agent-quota: a separate usage command today; a bus service is planned.
• Attachments: planned, not yet released.