Skip to main content
Smithers · closed alpha

You describe the work. Smithers does it.

An agent that already knows what is in your repos. It works in a real filesystem and git environment sitting above your GitHub, proposes the next thing worth doing, and does it where you can watch every step, stop it mid-sentence, and reverse anything it changed. GitHub stays your system of record.

Smithers is a closed alpha, open to design partners by invitation. There is no public sign-up yet. It runs in the browser, over GitHub, with nothing to install and no terminal required.

Nothing to assemble. You talk; Smithers creates the Flow that does the work.
Nothing to configure. No model to pick, no harness to wire, no clone, no local setup.
Nothing leaves your repos without an explicit yes from you, every single time.
smithers · your workspace github
smithers

I read your repos while you were signing in. You have 14 open issues across 3 repos. Four have been waiting more than a week.

Recommended
Why now: both pull requests have been green for nine days with no reviewer assigned.If you accept: I read both diffs and draft a review on each. Nothing is posted until you say so.
Accept ⏎Edit·Dismiss (one key, and it does not argue)
/ ask for anything, or press Enter to run the recommendation

An illustration of a first session. The content shown is an example, not a customer.

The first five minutes

Sign in, and the work has already started.

No blank prompt box, no setup wizard, no demo video. Signing in and connecting your first repos are one act, so by the time you can type, Smithers has read enough to be useful.

01 · Sign in with GitHub
Two clicks or fewer. Exactly what Smithers will read, and what it will never do without asking, in plain words before the permission screen.
02 · The workspace is already there
Your repos exist as real files with working git, in a hosted environment above your GitHub. Nothing to clone, nothing to install, no terminal. This is the product: a place the work happens, not a view of GitHub.
03 · Smithers opens with what it found
The first message is not a greeting and not a question. It is the state of your actual repos: what is open, what is stalled, what has been waiting on someone.
04 · One recommendation, not a menu
The best next thing, grounded in real evidence from your repos. It says what it proposes, why now, what happens if you accept, and how to accept, edit, or dismiss it. Dismiss is one key, and a dismissed recommendation does not come back unchanged.
05 · Accept, and it runs in front of you
Visible in the timeline while it works, and an embedded card in the conversation when it lands. Never a full-page takeover, never a silent completion, never an anonymous spinner.
06 · The result leads with the result
What you wanted comes first, in words. Then the state, or the one decision that needs you. No score, no grade, no dashboard to visit, and no “was this helpful?” ever.
07 · At most three questions
Smithers asks only what it genuinely cannot work out, like what counts as done for you. Never anything it could have read from your GitHub.

Usage is metered and shown in real dollars on the run itself. During the alpha, design partners never enter a card.

While it works

Watch it. Steer it. Take it back.

An agent you have to babysit is not saving you anything, and an agent you cannot see is not one you can trust. Work in motion stays in front of you, and every change it makes has a way back.

Anything that is not instant runs as visible work in the timeline, and it is still findable afterwards if you walked away. Nothing completes or fails silently.
Escape halts the work in front of you within a second and tells you what it stopped. Redirect it, narrow it, or take over yourself.
Edit its output, change its mind, or reverse what it did. Smithers confirms in one line what changes now, and the correction is the clearest thing you can tell it about what good looks like.
Every action has a keyboard path, from the first run onward. Press / for everything Smithers can do, with the recommendation first.
in the conversation working
you > go ahead, but let me see the review before it goes up
read both diffsdone
checked the failing test historydone
drafting the reviewworking
post it to GitHubneeds your yes
esc stops it · the exact comment is shown before anything is posted
Consequence-scaled approval

Nothing leaves the building without your yes.

The strength of the control scales with the consequence of the act, so routine work does not nag you and irreversible work cannot slip past you.

Acts freely
Anything reversible inside your own workspace. No prompt, and still fully visible in the timeline.
Asks once
Changing files in a scope you already connected. One approval, remembered for the session, revocable whenever you want.
Asks every time
Pushes, pull requests, issues, comments, deletions. An explicit yes each time, showing exactly what will happen. Never remembered, never assumed.
The rule under the rule

An approval covers what you were shown and nothing more. If the work grows past it, Smithers stops and asks a new, smaller question instead of proceeding. And it never claims to have done something it did not do: if the change did not reach GitHub, it says so.

What makes this different

There is nothing here for you to assemble.

Everything else in this category hands you an authoring surface and makes you the author: you lay the steps out, wire them together, and own them forever. Smithers has no authoring surface, and that is the product. You describe the work in the conversation, and Smithers creates the Flow that does it. There is no manual editor, and there will not be one. It is a product boundary, not a missing feature.

“Every Monday, tell me which pull requests are stuck and why.”

→ Smithers creates the Flow, shows you what it will do, and from then on it is yours to run, edit, or retire

“Take the four issues that have been sitting the longest and tell me what it would take to close them.”

→ it works in the real repo, in real git, and comes back with the answer and the one decision it needs from you

Everything Smithers can do is a Flow, and every Flow is reachable three equivalent ways: ask for it in the conversation, type its /name, or use the affordance on the work itself. No path is privileged, and there is no capability hiding on only one of them.

Open source underneath

The engine is public. You are not locked in.

Smithers runs on an MIT-licensed durable orchestration engine that anyone can read, run, and host. That is a reason to trust the product, not the pitch for it.

Your code, history, issues, and pull requests stay in GitHub, which remains your system of record. Smithers changes where you work, not where your work lives.
The engine that executes every step is open on GitHub. You can read exactly how a run is checkpointed, retried, and rolled back.
The engine runs on your own machine or your own infrastructure, with no cloud required. It drives the Claude or Codex subscription you already pay for; API keys are optional.
For developers · run the engine yourself today
$ bunx smthrs initcopied ✓

Installs the engine into your project with a runnable example you can execute immediately, plus the smithers skill for the coding agent you already use. Describe what you want and your agent creates it, with no configuration. This is the open-source route, and it is separate from the hosted product above.

Engine docs, written for agents →Projects built on it →Orchestration patterns field guide →

Stop babysitting the agent.

Smithers is for technical operators whose work already lives in GitHub: people running agents, maintaining repos, and carrying a queue of issues and pull requests in their head. If that is you, tell us about your repos.

Closed alpha, by invitation, with no public sign-up yet. We are onboarding design partners one conversation at a time.