Hermes Agent alternative: Muster, when the run has to be auditable.

Hermes Agent is a serious piece of engineering and the comparison should start by saying so. Seven terminal backends is not a marketing bullet — being able to run the same agent locally, in Docker, over SSH, in Singularity, on Modal, on Daytona, or on Vercel is real portability work, and most harnesses do not attempt it. The FTS5 cross-session search with LLM summarization is a better answer to "what did we talk about last month" than most projects have. And by adopting the agentskills.io open standard rather than inventing a private skill format, Hermes made a genuinely community-minded choice.

Muster is not going to out-build that. Anyone telling you a young project will out-feature 25,737 commits is selling something. The argument for Muster is a different axis entirely.

The question Hermes does not try to answer

Ask any harness — Hermes included — "which files did this run modify?" and it will answer from its own event log. That log is populated by the agent loop reporting its own actions. The reporting is honest; the coverage is not complete, because modern coding agents perform most edits by shelling out. A python - <<EOF block that rewrites four files is, from the structured-patch channel's point of view, one opaque shell command.

Muster measured how bad this gets rather than speculating. Against the Codex app-server protocol, across five live runs — three protocol probes plus two full head-to-head recordings — the structured file-change notification fired zero times out of five while all the edits landed on disk. Both of the head-to-head runs also showed zero turn/diff/updated notifications. The self-report channel was not merely lossy. It was empty.

Muster's workspace observer sidesteps the whole reporting question. It pins a git baseline at start(), then re-derives the entire change set from git plus the filesystem on every cycle, hashing file contents before and after. Filesystem watch events are used only as a signal to re-derive — never as data — because on macOS they arrive type-less, coalesced, and sometimes with a null filename. It is strictly read-only against your repository and its git index: never git add, never the index lock, with all writes confined to a shadow tree outside the watched root so the observer cannot trigger itself. In the two head-to-head runs it reported the change in 86ms and 75ms against a 1000ms budget, with a git apply-verified diff.

Memory: search versus scope

This is the sharpest genuine difference and it is worth being precise about. Hermes's cross-session recall is a search problem well solved: index everything, retrieve by relevance, summarize. Muster's is a boundary problem: a memory belongs to a tenant, a workspace, a user, a role, or a session, and retrieval is filtered by which of those lanes the current run is permitted to see, with tests that specifically try to make one scope leak into another.

If you are one person running one agent over your own work, search is straightforwardly better — it finds more. If you are running agents for several customers on shared infrastructure, unbounded search across all your sessions is precisely the thing you must not have. The two designs are optimising for different rooms.

Skills: autonomous versus gated

Hermes lets an agent create its own skills. Muster deliberately will not, and treats that restraint as a feature: feedback has to become a test fixture before it is allowed to change durable behaviour. That is slower and less impressive in a demo. It is also the only version of "the agent learned" that you can put in front of a reviewer, because every behaviour change has a test attached explaining what it is supposed to do and a way to check it still does it.

Choose Hermes Agent when

• Execution portability matters — you need the same agent on Docker, SSH, Modal, Daytona or Vercel.
• You want the largest ready-made tool library and an open, shared skill standard.
• Broad recall across your own history is more valuable than strict separation between contexts.

Choose Muster when

• You will have to demonstrate which files an agent changed, to someone who will not accept the agent's own word for it.
• Multiple tenants, customers, or teams share one deployment and memory must not cross between them.
• You want token spend and replay waste in your own ledger.
• Frappe or ERPNext is the system of record — Muster's depth there is real, while its general breadth is not.

These are not mutually exclusive positions. Muster is designed to wrap other runtimes rather than replace them, and treats Codex, Claude Code, Cursor SDK, OpenHands and Pi as adapters behind its own contract. Governance staying outside the backend is the whole point.

Compare against OpenClaw → Compare against QM → Read the docs →