Hermes Agent, from Nous Research, is one of several coding agents that can now join a shared workspace as a member rather than a bot: MarkTechPost reported three integration paths into Block's Buzz, the open-source Nostr workspace where agents carry their own keys. If you run Hermes, that is worth knowing. So is the alternative: keep it on your machine and put a room around it, so it behaves like every other agent your team can reach.
Why the runtime should be the boring part
An agent is a runtime pinned to a working directory. Which runtime it is decides how good the output is; it should decide nothing about who may ask for that output, what pauses for a human, or what the record looks like afterwards.
That is the design muti settled on. A muti agent is Claude Code, Codex, Hermes, OpenCode or Cursor plus a directory and a permission mode. Whichever one answers, a request gets the same public id, the same live timeline of tool calls and diff summaries, the same per-person permission, the same four action types that always stop for the owner, and the same audit line. Swapping the runtime behind an agent changes the quality of the work and nothing about the governance around it.
Any runtime on the machine; one room with the same requests, approvals and audit line.
The two shapes for Hermes
Neither is wrong. If your team already runs a relay, a member is the natural shape. If Hermes is configured on one laptop and three colleagues need things from it, a room is the smaller move. The wider comparison is in Block's Buzz and muti.
What changes when the runtime is not Claude Code
Three practical things, worth knowing before you promise a teammate anything.
Detection. The daemon looks for what is installed on the machine and reports it with versions, so an agent can only be created on a runtime that is actually there.
Session continuity. Each runtime exposes its own session reference, which muti stores with the request so an interrupted run resumes under the same id rather than replaying from scratch. See request ids and resumable runs.
Permission vocabulary. Claude Code has plan, default, acceptEdits and bypassPermissions; other runtimes have their own ladder from suggest-only to full auto. muti maps each to the nearest level and then layers per-person grants on top, so the four always-pause categories behave identically. See permission modes, explained.
Running more than one
A room can hold several agents on different runtimes, and people pick the one they want by name. A common setup is one agent per repository rather than one per model, named for the job: pipeline, site, notes. Each has its own serial queue, so a long Hermes job does not block a quick Codex question, and a request can reference another agent's earlier result by id. See Codex and Claude Code in the same room.
Frequently asked questions
Is Hermes supported in muti today?
The daemon detects Hermes on a machine alongside Claude Code and Codex, which are the runtimes with drivers verified end to end. Cursor is implemented and awaiting a login for verification; OpenCode and DeepSeek are in progress. The room model is identical for all of them.
Does a room work with a self-hosted model?
The room never sees the model. It hands a request to the runtime on the machine, and the runtime decides what it talks to, which can be a local model. Nothing about the approval flow changes.
Can one agent be in a Buzz channel and a muti room at the same time?
Yes. Both use outbound connections from the machine and neither opens a port. Which one holds the approval for a given request is decided by which one the request came through.
Which runtime should I pick?
Pick per job, not per team. The point of making the runtime boring is that you can change your mind later without re-explaining the rules to anyone who uses the agent.
Sources
- Nous Research and Buzz, MarkTechPost: marktechpost.com
- Buzz: github.com/block/buzz
- The category map: where agents live in 2026
Whatever runs behind it, give your agent a room.
Опубликовано