Rent the staff. Own the brain.
Features waiting for capacity? RepoStaff is being built to extend your software team: agents will take scoped work from the backlog to a reviewable PR, with lasting context and oversight. Your team sets the pace.
decisions/why-this-exists.md
The backlog has plans. Give it capacity.
You have features worth building and a team with finite capacity. RepoStaff is being built to augment that team with task-based execution, persistent context, and configured review.
The aim is more of your backlog within reach, without making every ticket another coordination job for your seniors. The knowledge will accumulate in a repo you own.
Hypothesis, not traction. Written for engineering CTOs and senior engineers. It comes from one engineer's work and hasn't been tested with other teams.
org/staff.md
Three pillars you rent.
Ephemeral agents
Prove it. Then disappear.
One fresh VM per task, with Docker and a browser. Each will implement, use your app, attach evidence to the pull request, and get torn down. No production credentials.
Long-term agents
Marshall stays. One per org.
The planned staff-engineer job: interview you, plan and triage tickets, talk in chat, and check every ephemeral agent's work. Don't like the name? Rename it.
Management UI
The receipt.
Oversight of every run and gate, your connections and OAuth, and cost against outcomes. Estimates will say they're estimates. If the agents are wasting your money, you should see it here first.
The brain feeds all three. It isn't rented. It's a git repo in your org.
memories/ownership.md
The brain is a repo.
Every memory will record its source. Agents will get a pinned, read-only copy. When a person and Marshall disagree, the person's change wins.
Your exit, any day you want it
$ git clone git@github.com:your-org/brain.gitorg/how your company worksdecisions/why things are the way they arememories/what Marshall learned, and whereskills/what your agents know how to dopolicy/who has to approve what
Cancel and you keep every piece of knowledge committed to the brain. You lose the Management UI and the orchestration, so export run history and session archives separately; credentials and live services require separate migration.
README.md#getting-started
Five commits to ready.
The planned order. None of this runs yet.
Select repos and a brain.
Install the GitHub App on the repositories you pick. Create a brain repo or connect one you already have.
Add your key.
An Anthropic or OpenAI API key, on your own account. No subscription logins. No resale.
Enroll compute, tracker, chat.
Enroll a runner on your compute, then connect one tracker and one chat.
Marshall reads. You confirm.
With the read access you authorize, Marshall reads, then interviews you on the gaps. What you confirm lands in the brain with its source.
Label it ready.
A configured, authorized ready label or board column will start the run.
policy/workflow.yaml
You set the gates.
Cautious. The planned default. A person approves the plan, approves the pull request, and merges it.
- triggeredAgent
- plannedHuman
- implementedNone
- validatedCIthenAgent
- reviewedAgentthenHuman
- mergedHuman
Tests. CI. Marshall. Your approvals.
The evidence on the pull request will support your review. It won't guarantee the code is correct.
Merge isn't deploy.
RepoStaff will stop at merge. Agents get no production credentials.
Policy lives in the brain.
Gates and limits will sit in
policy/workflow.yaml, in your repo.The Management UI changes it by commit.
The record of who loosened what will be
git log.Agents read a pinned snapshot.
Ephemeral agents will get a read-only brain snapshot. They can't write policy or grant privileges.
Chat proposes. It doesn't authorize.
A request in Discord or Slack is a proposal. It will never authorize execution by itself.
decisions/where-it-runs.md
Hosted management. Your compute.
Management UI and control plane
Hosted by me first, to prove the concept. Compute comes separately, in this order.
- Full self-hosting
- Management self-hosting is demand-led, post-MVP work. No date. No setup or support promise.
- Source access
- Paying customers will get source and an issues and PR path. Exact license deferred. Not open source.
- Your data
- Agents on your hardware don't keep your code on your network. Your model provider gets context. No data residency guarantee.
skills/connectors.md
Merge order.
Every connector is planned. None of them exist yet.
First
Planned firstGitHub Code, CI, Issues, and the brain repo. The only code host and CI at launch.
Discord The first chat.
Next
Planned nextLinear Tracker.
Slack Chat.
Then
Planned to followTrello Tracker.
Stretch
Stretch goal, not promisedMicrosoft Teams Chat.
Models. Your Anthropic or OpenAI API key. No subscription logins. I won't resell tokens or mark them up.
Harnesses. Hermes as the Marshall runtime. OpenCode first, then Claude Code headless, for ephemeral agents. All on trial, all planned. These are harnesses, not model providers.
README.md#pricing
Flat per org. No seats.
Planned prices are untested hypotheses. They may change, and nobody has tested them with a buyer yet. Model usage goes to your provider on your own key, and self-run compute is your bill. Budget controls are planned and unproven, and no RepoStaff cap limits your provider's invoice.
More work when you need it. No task-usage charge for compute you do not consume when you pause. The subscription, persistent Marshall, and shared infrastructure overhead continue between tasks. Hosted compute will include 100 VM-hours, then metered usage at the planned service rate. Hosted usage pricing includes a service margin; final commercial terms remain hypotheses. Model tokens stay on your separate BYOK invoice.
Design partner
$149
Half the self-run list price, or $349 if you choose hosted once it exists. Billing starts when the first agent run completes on your repo. Cancel any time. A refund for a month when no pull request merged is planned, but its exact conditions aren't settled.
Become a design partnerSelf-run
$299
Hosted Management UI, your compute. Marshall, ephemeral agents, and the brain repo included. No limit on members or repos.
Become a design partnerHosted
$699
Hosted Management UI, my AWS. 100 VM-hours included, then $0.60 per VM-hour up to your cap. Opens only after the paid self-run launch.
Hosted compute waitlist detailsA $49 self-hoster rate is planned for up to three members on their own hardware, still on the hosted Management UI. No sooner than 90 days after the paid launch, and only if the support load allows. No customer AWS connection, email support or trial. Self-hoster rate details
Before you ask.
Is this available now?
No. RepoStaff is being built, and everything on this page is planned. Design partners come first.
Can I self-host the whole thing?
Not at launch. Marshall and the ephemeral agents will run on compute you pick, and the Management UI starts hosted by me. Full management self-hosting is demand-led, post-MVP work, with no date and no setup or support promise.
Is it open source?
No. Paying customers will get source access and an issues and PR path. The exact license and contribution terms are deferred. The brain is your own repo, so that's yours regardless.
Where will my code go?
Running agents on your hardware doesn't keep your code on your network. Selected context goes to your model provider, and source may pass through the hosted control plane. Launch customer scope is US-only. US primary hosting is intended, exact regions aren't set, and there's no data residency guarantee.
Where do my model keys go?
To RepoStaff. By default they'll go into an encrypted vault, and agents will reach your provider through a broker with a short-lived token per job. Your provider still receives task and code context, so bringing your own key isn't a no-retention promise. None of this is built yet.
Will Marshall read all our channels?
No. It will work in DMs and in threads it's mentioned into. Discord and Slack will still send RepoStaff other channel events, and RepoStaff will discard those without storing them or passing them to Marshall.
Is Marshall as good as a staff engineer?
I don't know yet, and neither does anyone selling you an agent. Marshall has that job. The Management UI will show you how it's going.
What happens to the brain if I cancel?
Nothing. It's your repo. You keep every piece of knowledge committed to it, and you lose the Management UI and the orchestration. Run history and Marshall's session archive are separate exports, so take those first.
Should I just buy Devin?
Maybe. My unverified market hypothesis as of 2026-11-08: it was the only product I found that seemed to deliver all three pillars, it's further along than RepoStaff will be for a long time, running it in your own VPC takes their Enterprise plan, and what it learns lives inside their product. If neither bothers you, go buy it.
Do you have SOC 2?
No, and it isn't in the first version. There's no penetration test or completed audit either. If you need those today, I'm not ready for you.
Harry Thomas, founder and, for now, the whole company, is looking for
Five teams. At most.
Audience hypothesis — fit: Five to thirty engineers. Code on GitHub. Already using AI coding tools. Seniors are the bottleneck. You care where your code runs. Your lead can pay by card.
Not yet: Regulated enterprise. GitLab or Azure DevOps. Teams on Jira. Management on your own infrastructure. Teams outside the US, at launch. Anyone who needs a data residency guarantee.