A self-hosted AI agent that runs entirely on your own infrastructure.

What "self-hosted" actually covers here

Self-hosted is an overloaded word in agent tooling. Plenty of products describe themselves that way while the memory index, the usage accounting, or the audit log still lives on someone else's server. The useful question is which parts of the run leave your machine, so it is worth being specific about Muster.

• The session store is local. Sessions, messages, recalled memory, and usage rows sit in a SQLite database on the machine running the harness. There is no account to create and no telemetry requirement.
• The token ledger is yours. Every run records tokens in and out, estimated cost, and a replay-waste flag locally. You are not reading your own spend back out of a provider dashboard.
• The model is a route, not the platform. Muster routes to cloud APIs, aggregators, open-weight models, or self-hosted endpoints. Governance lives in the harness, so switching providers does not mean rebuilding memory scoping, policy, or evals.
• The workspace is read-only to the observer. The file-change observer never runs git add, never writes objects, and never takes the index lock. Its only writes go to a shadow tree outside the watched root.
• Channels terminate on your gateway. Terminal, web, Telegram, Slack, Google Chat, Teams, Discord, and WhatsApp all enter the same envelope on infrastructure you run.

Why teams self-host an agent in the first place

Usually one of three reasons, and they want different things from the harness. The first is data residency: the work touches customer records, source code, or an ERP system that contractually cannot be shipped to a third-party agent service. For those teams the deciding features are scoped memory boundaries and the ability to point every model call at an endpoint inside their own network.

The second is cost control. A long-running agent that replays its entire history every turn burns money in a way that is invisible until the invoice arrives. Muster measures this deterministically: across five scenarios and 170 turns, a replay-everything harness sends 875.8k tokens where Muster sends 355.2k — a 59.4% reduction, and 62.7% on the longest thread. No model is called during the benchmark; run muster benchmark and you will get the same numbers.

The third is audit. Someone downstream — a security reviewer, a customer, a regulator — will eventually ask which files an agent modified and how you know. Most harnesses answer from the agent's own report of its actions, which is a problem, because coding agents do most of their editing through the shell where the structured patch channel never fires. Measured live against Codex across five runs, that channel reported zero of the real edits. Muster's observer derives the change set from git and the filesystem instead, catching the edit in 86ms with a diff verified by git apply and a receipt hash that reproduces on a machine that was never there.

What it does not give you

Muster is early and deliberately narrow. It ships 34 capability packs of which only two are substantial; it is not a plugin ecosystem and will not become one. If you need the largest available library of ready-made integrations, a larger project is the better answer and this page is not going to pretend otherwise. What Muster is trying to be is the layer that makes a self-hosted agent defensible after the demo.

See the observed file-change evidence -> Compare with OpenClaw -> Read the docs ->