What is the Monitor Protocol?
The Monitor Protocol is the Claude Code Monitor tool idea, taken out of the session and made durable, shareable and accountable.
The Monitor tool has the right instinct. You point it at something, it watches, and it tells you when something happens. But it stops when the session stops, it runs for thirty minutes at most, only the session that started it can see it, and it remembers nothing. So every session starts its own watches again, and whatever happened in between is gone.
A monitor in this protocol keeps the instinct and drops those limits. It exists on its own. Many people and agents can subscribe to it. It remembers what it saw, says who saw it and how much to trust it, and keeps a running summary that says out loud what it does not know. If a subscriber is away, nothing is lost. When it comes back, everything it missed is waiting, in order.
Think of it like a mailbox with a doorbell. The doorbell tells you something arrived, but you might not hear it. The mail stays in the box until you take it out. In the Monitor Protocol, notifications are the doorbell and the log is the mailbox, so a missed notification never means a missed event.
What can the Monitor Protocol enable?
- A Claude Code session can follow one monitor for all of your builds, pick up where it left off after a restart, and never handle the same event twice.
- A team dashboard and an agent can subscribe to the same monitor, each with its own filter and its own place in the log.
- An agent can act for you and say so. The record keeps the agent's name and yours, so anyone reading later can tell them apart.
- Two workers can share a stream of alerts without doing the same job twice, because a lease gives each piece of work to one of them at a time.
Why does the Monitor Protocol matter?
Depending on where you sit, the Monitor Protocol helps in different ways.
- Developers: one small protocol for watching things, instead of a new polling loop, webhook handler and state file for every source.
- AI applications and agents: a durable way to be told about the world. Sessions can come and go. The monitor keeps watching and keeps each reader's place.
- People: a record you can check. Every observation says who made it and how much weight it can carry, and a machine cannot give itself the top rating.
How it relates to MCP
The Model Context Protocol lets an AI application reach out to the world: read a resource, call a tool. The Monitor Protocol covers the other direction. It is how the world reaches back, reliably, over days instead of seconds.
The two fit together. A Monitor Protocol server can also run as an MCP server, so any MCP client, including Claude Code, can create monitors, publish, subscribe and read observations as ordinary tools. See Claude Code and MCP.
Choose between a monitor and a direct tool call
| Pattern | Best suited for |
|---|---|
| A direct tool call | Asking once: "What is the status of build 412?" |
| A watcher inside one session | Following one thing while you work, when losing it at the end of the session is fine |
| A monitor | Following something for days, for more than one reader, with a record that survives restarts and says how much to trust each entry |
Built on standards you already know
The Monitor Protocol does not invent a new envelope, transport or delivery model. Observations are CloudEvents. Filters use the shape of Nostr filters. Cursors and acknowledgements follow Kafka's consumer offsets, and leases follow Kubernetes leases. What the protocol adds is a small set of rules about evidence: who saw it, how much it can be trusted, and what is still unknown.
Where it runs today
- The reference server and command line tool, published on npm as
@mentu/monitor-protocol. - Any MCP client, through the server's MCP mode.
- Atrio, the event log we use to run our own work, which serves the protocol at
/mp/v0and runs the same conformance suite.
The Monitor Protocol is at version 0.1. The specification, schemas and conformance suite are public on GitHub under the Apache 2.0 license, and details may change before version 1.0.