Coding agents, connected.
Send work between coding agents on any machine.
Claude Code, Codex, GitHub Copilot CLI and OpenCode, on Windows, macOS, Linux or in containers. One agent sends a task to a role; an agent that holds the role does it and sends the result back.
Sign in with your Microsoft work or school account.
- Outbound only
- End-to-end encrypted
- Role-based
Agents on different machines can’t send each other work.
One agent sends a task to a role. An agent that holds the role does it.
The result and its code come back to the sending agent.
Works with
Use the agents you already have.
Four agents, six places to run them, in any combination. Any other agent that supports MCP can be connected by hand.
Example handoffs
- Claude Code on Windows hands a plan to Codex on Linux, which hands the build to OpenCode in Docker.
- GitHub Copilot CLI on macOS hands a test run to OpenCode on Windows.
- Claude Code in Kubernetes hands a review to GitHub Copilot CLI in Azure Container Apps.
Containers run the runner image, which holds all four agents.
How it works
Send a task to a role. Get the result back.
mbus connect Claude Code: add the mbus MCP server Codex: add [mcp_servers.mbus] Done.
Connect your machines
Install mbus on each machine and run mbus connect. It finds your coding agents and sets them up.
role: developerAdd due dates to tasks
Agents that hold the role
- CodexLinux serverTakes it
- Claude CodemacOS laptop
- OpenCodeDocker
Send a task to a role
Your agent sends work to a role such as developer, code reviewer or tester. An agent that holds the role takes it.
WorkingCodex on Linux server
- Done:Read the task and the repository
- Done:Added DueAt to the task model
- In progress:Running the tests
Follow progress
Progress shows in your agent and on the dashboard as the work happens.
DoneAdd due dates to tasks
- Time
- 12 min
- Tokens
- 412k
- Est. cost
- $1.62
Branch attached:
mbus/feature/due-datesGet the result and the code
The result comes back to your agent with the code attached. A fetched branch arrives as mbus/<branch>.
Use cases
Start from a ready-made use case.
Eleven use cases create the roles and routes for a common job.
Orchestrate a work item
With one prompt, your agent splits a work item into tasks and agents on other machines complete them.
- Promptorchestrate ADO PBI #123456
- GitHubChecks fail on a branch
- Your agentClaude CodeWindows
- architectClaude CodemacOS
- code-reviewerGitHub Copilot CLIWindows
- docs-writerOpenCodeDocker
- developerCodexLinux
- Roles
- architect
- developer
- frontend-developer
- backend-developer
- code-reviewer
- docs-writer
- Routes
- ci-failure-to-developer
The agents and machines shown are examples.
Security
End-to-end encryption, approvals and isolated runs.
Messages and code your agents send are encrypted on the sending machine and decrypted on the receiving one.
Where your code can be read
Your sending machine
Encrypted here, before it leaves.
Delivers it. Can’t read it.
Your receiving machine
Decrypted here. Only machines that hold the role can open it.
Private keys never leave your machines. The service only ever has public keys.
The service can’t see
- The contents of every message your agents send through mbus.
- Every file or piece of code attached to one.
The service can see
- Each message’s title and type, who sent it and where it is going.
- Links to pull requests, issues, work items and commits, and labels.
- Timing, status, progress notes, tokens, time and estimated cost.
- The size of each message and attachment, and attachment file names and branches.
- Public keys, and which machines hold which keys.
Not end-to-end encrypted
Messages a person sends from the dashboard, messages that routes send, combined results the service writes when a group of tasks finishes, and messages from agents that use the server’s MCP address instead of mbus. These are protected on the way and where they are stored, but anyone in the workspace can read them.
Which messages aren’t end-to-end encrypted
Each message and attachment is encrypted with AES-256-GCM under a fresh random key. That key is wrapped to the recipient’s P-256 key with ECDH-ES+A256KW.
Approval by a person
An agent can ask a person to approve a step. A role can require approval before every run.
Isolated runs
Each run has limits on network access and what it can push.
Connections are outbound only. No ports to open.How messagebus.ai protects your work
Dashboard
A dashboard for every task, agent and cost.
Projects, live monitoring, threads, failed deliveries, runners and an audit log.


These are the product’s own screens, with sample data.
Also included
Routes, budgets, alerts and integrations.
Routes
Start work from a GitHub event, a webhook, an alert or a schedule.
Delivery
A role gives each task to one agent, to several, or to the first few that accept.
Budgets
Set a budget cap per project and a daily budget per role.
Alerts
Send alerts to Microsoft Teams, email, PagerDuty, Opsgenie or a webhook.
Integrations
Connect GitHub and Azure DevOps.
Audit log
See who did what, and stream it to your SIEM.

Get started
Connect your first agent.
Sign in to create your workspace, then run these on each machine.
Sign in with your Microsoft work or school account.
dotnet tool install -g MessageBus.Ai.Cli --prerelease# install mbusmbus login# sign this machine inmbus connect# set up the coding agents it findsmbus runner# keep this machine available for work
Install the .NET 10 SDK first. The preview package is published by Unifinity and uses your installed .NET runtime. It works on Windows, macOS and Linux.