Home / agent-bus
🏢 An enterprise platform that connects your AI agents, tools and people, governs every access, and runs each agent safely in its own box.
💻 Runs on Linux and macOS: everything on Linux; on macOS the bus, the runner and the command line.
Not a single tool but a whole platform: it runs on localhost and grows to many hosts, users and teams.
What it gives you
- 🔗 Connect: reach any assistant, script, service or person by name.
- 🔍 Discover: see who is there and what each one can do.
- 🛡️ Access Control: decide who may call, read and manage each name.
- 🔑 Secrets: shared passwords on the bus; a script's own stay on its host.
- 🔒 Secure communications: other machines can connect over HTTPS with a pinned certificate, or over SSH; one port serves HTTP and HTTPS; on one host, a private socket only your account can reach.
- 📦 Confinement: scripts run walled off, with memory and process caps.
- 🧱 Ready helpers: chat, mail, text messages, databases, server checks.
- ⏰ Scheduling: wake an assistant at a time or on a rhythm.
- 🚨 Alerts: a server error reaches an assistant, then you on Slack (or Telegram, SMS).
- 🌐 Clients: libraries for six common programming languages.
- 🖥️ Web console: see and manage everything from one page.
- 📥 Install: one command, a ready-made container or a package.
- 🧠 Search by meaning: your AI finds and sorts your documents.
Why you want it
- 🤝 Your AI assistants work alone. Now they message each other and you by name. How
- 📬 Someone is away. A message waits in their inbox until they read it. How
- 🔑 You cannot trust an assistant with a password. Now it uses your tools through helpers that hold the passwords. How
- ⏰ Routine work waits for somebody to remember. Now reminders and daily checks wake an assistant by themselves. How
- 🚨 You hear about breakage late. Now a server problem reaches an assistant that tells you on Slack (or Telegram, SMS). How
- 🔄 Updating servers means logging in to each. Now one code push updates them all and reports back. How
- 🧰 Your handy scripts stay private hacks. Now any script becomes a safe, shared team tool. How
- 👀 You cannot see what is going on. Now one web page shows what needs attention. How
The bus
Everyone finds each other; you decide who can do what.
The problem: AI sessions, scripts and people cannot find or reach each other, and nobody controls who may use what.
- 🔗 Connect: reach anyone by name, on this machine or another.
- 🔍 Discover: see who is there and ask each one what it can do.
- 🛡️ Access Control: for each person and each agent, decide who may call it, read it and manage it.
- 🗂️ One place for settings and shared passwords: kept with each name, readable only by its owner, its managers and the agent itself.
More: access.
🔐 Passwords in two places: the ones several people or machines share live on the bus; a script's own stay with its runner, on its host. Two stores
The runner
systemd on steroids, for AI-era tools. The problem: a useful script is unsafe to share, or a pain to run as a service.
- 📦 Confined: by default it sees only its own folder and cannot touch its neighbours.
- 📏 Capped: memory and processes are limited; a host that cannot confine it refuses to run it.
- 🔑 Its own passwords: kept in a file on its host; the bus never shows them, and set on the host they never cross it.
- 🎛️ Controlled from a web page: start, stop and restart it, and see its state and why it stopped.
- 🔁 Kept running: a helper that stays running is restarted when it fails; a script run per message starts fresh each time.
Under the hood: cgroups, a namespace sandbox with a system-call filter, and the runner's own unprivileged account. More: isolation.
Helpers are building blocks
🧱 Our agents are building blocks, like lego: they make it possible to assemble complex systems easily, without sacrificing security.
- 💬 Chat: Slack and Telegram.
- ✉️ Email: read, write, connect a mailbox to an agent or queue.
- 📱 Phone: a text message.
- ⏰ Schedule: at a time, on a rhythm, or later.
- 🖥️ Servers: logs, disks, services, code.
- 🗄️ Data: safe database access for your AI, from read-only to full access — the password never leaves the server.
- 🧠 RAG: a shared memory for your AI: it sorts your texts into your own topics and finds them by meaning.
- 🌐 Web: AI reads web pages as text or Markdown and makes bounded HTTP requests; limit it to the sites you allow.
- 🤖 AI: ask a model.
- 🪟 Sessions: start whole fleets of your sessions with one call — AI chats, monitors, anything in a terminal — and bring them all back after a reboot the same way.
A mail help desk, snapped together from four blocks:
🔒 Why it stays safe: each block gets only the access you grant, and a block that holds a password keeps it: your AI uses the block and never sees that password. Help desk
From one laptop to many hosts
- 💻 Start small: bus, runner and AI sessions on one laptop.
- 🖧 Add servers: each runs its own runner and helpers against the same bus.
- 👥 Add people and teams: users, groups, and who may use what.
One real flow: a server error reaches you on Slack
- 🧩 Every step is a block above: the log watcher, server checks and Slack. Log watcher
Eight things you can do
- 🔍 Ask a second assistant to review the first one's work, and wait for the verdict.
- ⏰ Every weekday at nine, wake an assistant to review the overnight server logs.
- 🚨 When a server logs an error, an assistant looks into it and tells you on Slack (or Telegram, SMS) what it found.
- 📜 Get a daily Slack note about certificates that expire soon.
- 🔄 Push your code once and every server that follows it updates itself.
- 🗄️ Let an assistant answer "how many orders failed yesterday" without ever holding the database password.
- ✉️ Have an assistant read incoming support mail, draft the reply and send it after your approval.
- 🌙 Send a task to an assistant on another machine tonight and read its answer in the morning.
Start here
- User guide: the shortest path from nothing to a running assistant or script.
- Install, details in setup.
- Source and docs: github.com/parf/ai-agent-bus.
Details
Made for agents
Web management panel
Integrations
Tools built for agent-bus — suggested ways to put it to work.
mdhouse — chat with an agent about a document
mdhouse is a local, git-aware Markdown workspace. With
mdhouse --agent-bus it gains a chat panel: ask a running Claude Code, Codex or
OpenCode session about the page you are reading. The agent edits the file, the page updates
live, and each page keeps its own history.
No adapter on the agent side: the panel sends ordinary bus messages and the agent replies as to any peer. Works with a local or a remote bus — mdhouse Agent Bus integration.
Plans
License
PolyForm Noncommercial; the client libraries are MIT.