It watches everything you build — your systems and your intent — without a secret ever crossing the wire. Like a ghost OpenClaw, but hyper-optimised for your project. No always-on computer needed.
# Runs on your machine — or your own GitHub (Actions) when it's off. Nothing to host. It opens branches; never merges.
Watching is cheap and constant; thinking is rare and aimed. That split is what lets it keep its eye on all your work for almost nothing.
# A model reads your repo exactly once — at bootstrap, for a few cents, it learns the shape of your logs and a map of your code. Keeping that current is free: the map rebuilds itself in deterministic local code whenever you commit, so a model never re-reads your repo again.
# So watching itself is plain deterministic reading — logs, commits, deploys, deduped against what it's seen. The model is the scalpel: it wakes only for a genuinely new signal, investigates it once, and remembers it, so the same issue never costs you twice.
It grounds your logs in your deploy and commit history, collapsing the noise into the handful of issues that matter — each a growing dossier.
It reads your coding sessions, distilling each into the intent you set and the blockers you hit — the thread no single session keeps.
It proposes the loops; you approve the ones you want; it works them on InferRoute's cheap off-peak and right-sized intelligence.
# As with all Steward work, your laptop can stay closed — GitHub (Actions) runs the loops, and a free InferRoute server ties it all together so you stand up no cloud of your own. Your code and secrets stay yours throughout. Whatever it builds comes back as a branch you review, CI-verified: it never touches your tree, never merges.
Its job is a defensible diagnosis — cause, evidence, options. What it can prove in your repo, it fixes and hands you the branch; what only production can confirm, it patches and watches; the rest — config drift, a design choice, a capacity or vendor limit — it lays out and leaves to you.
Your prompts never leave your machine — only compressed intent syncs, secrets scrubbed out first, so the cloud only ever sees redacted text, never your raw prompts, code, or keys.
* The redaction is deterministic and auditable — you can read exactly what's removed, and it runs only on your side.
** That blind coordination server is what lets your GitHub Actions deliver continuous intelligence — with no extra cloud service for you to stand up.
Inference inside a sealed enclave — where even we won't be able to read your data. Verifiable hardware, not a privacy policy.
Handling a big system isn't about watching more of it — it's about knowing its shape. The Steward is engineered for the two shapes most tools miss, and maps each deliberately.
Distributed deployments. It pulls a system's logs from across machines, services, and platforms — journald, files, SSH, Docker, Kubernetes, Vercel — into one ledger, so a fault anywhere is a single signal, not a hunt across dashboards.
Dependency-linked repos. It reads your manifests and works out which of your repos are first-party dependencies — so when a fault surfaces in your service but lives in a library, it opens the fix in that library, not the wrong repo.
# Most relevant once your system spans more than one host or repo — and it stays out of the way until then.
No more asking an AI what to do — then telling it to do it.
Point it at your work once, and it watches your systems, captures your intent, and helps design and maintain the loops it runs. Then it surfaces only what's genuinely yours — a one-click yes from your phone, or a real session when something deserves it.
On InferRoute's economy lane, tuned to spend only on what's new, it runs for pennies a day per project — and you stop wondering what's failing while you're away.
# Keep using Claude for what it's best at — powering through new features — while the Steward minds the rest.
Run it once — it sets itself up and starts the watch.
> The Steward — by InferRoute