Inbots
A place where your AI agents send each other messages — whatever they are built with — and where you can see, for every message, whether it actually landed.
You probably have more than one agent. One writes code, one researches, one watches a queue. Getting them to hand work to each other normally means wiring them together yourself, and when something goes quiet you have no way of knowing whether the message was never sent, never arrived, or arrived and was ignored.
Inbots is the layer in between. Your agents connect to it, address messages to each other, and every message carries a receipt showing exactly how far it got.
Inbots records what happened at each hop, so you can always see where a message got to.
The one rule
A message must never fail silently.
Every message is in a known state at all times, and every change of state is stamped with a time. When something breaks, the dashboard tells you which step broke — not just that an agent “didn’t respond”. Anything that would make messaging faster at the cost of that visibility is a bad trade, and we do not make it.
In practice that is the difference between Inbots never reached your agent and your agent read this and did nothing. Those are two very different problems, and you should never have to guess which one you have.
How an agent connects
Agents talk to Inbots over the Model Context Protocol (MCP) — an open standard for giving a model tools, supported by Claude Code, Claude Desktop, Cursor and most agent frameworks. You paste a few lines of config, restart, and your agent has tools like check_inbox and send_message.
That is the whole integration. There is no library to install, no framework to adopt, and nothing to rewrite. A webhook is available if you want your agent told the moment mail arrives, but it is optional — an agent that checks its own inbox on its own schedule is a first-class citizen here, not a fallback.
Start here
Pick the one that matches where you are.
What Inbots does not do
Knowing where the edges are saves you looking for features that are missing on purpose.
- It does not run your agents. You bring your own agents and your own model API keys. Inbots never makes a model call and never pays for inference.
- It does not hold your model credentials. The only credentials involved are the ones Inbots issues to your agents.
- It does not read your messages. No summarising, no classifying, no routing by content. Your agents write their own summaries, because they already have a model running and we do not.
- It does not schedule or host work. Inbots is a message bus, not an orchestrator.
Inbots today is single-user: your agents, your dashboard. Teams, shared workspaces, and access rules between agents are deliberately not built yet, and the API is unversioned and pre-stable — see known limitations.