Pillar 1
Ephemeral agents
Prove it. Then disappear.
- boot
- build
- prove
- gone
One task, one fresh VM. No production credentials.
Coming soon · AI staff for the team you already have
Hand it a backlog ticket. It will hand back a pull request with proof attached. You merge.
How it will work
You
RepoStaff
You
Example ticket · illustration
You step in 3 times. The default. A person approves the plan, approves the pull request, and merges it.
You
Start the ticket
Add the ready label to the issue.
Marshall
Check it's ready
Not a duplicate, right repo, enough detail to start.
Marshall
Write the plan
Plan and acceptance criteria go on the ticket.
Worker
Build it
Fresh VM. Branch, commits, draft PR.
Worker
Prove it
Runs your checks, attaches proof to the PR.
Marshall, then you
Review the PR
Marshall reviews before you do.
You
Merge
Ticket closes. Learnings are recorded.
Every preset: a fresh VM per ticket, CI before merge, and CI config, infra, migrations and .env files off limits.
Boss rule
RepoStaff stops at merge. Agents get no production credentials.
The proof on a PR supports your review. It doesn't guarantee the code is right.
House rules
policy/workflow.yaml.git log.Mission brief
Built by a staff engineer, for software teams.
The situation
Worthwhile features, finite capacity. What the AI staff learns lands in a repo you own.
Built for
5–100 engineers on GitHub, already using AI coding tools, whose seniors are stretched thin.
Roster
Pillar 1
Prove it. Then disappear.
One task, one fresh VM. No production credentials.
Pillar 2
Marshall stays. One per org.
Tasks come and go. Marshall stays across all of them and keeps what it learns.
Plans, triages and checks every task.
Pillar 3
The receipt.
Illustration. Stop or redirect any run.
mainThe brain feeds all three. It isn't rented. It's a git repo in your org.
Incoming
Marshall will take the first pass, so your seniors only open what's worth their time.
Yes, an AI staff company is offering to clean up after AI. I see the irony.
[in]5 new · illustration
Ready
Needs more
Duplicate
Needs a person
01 · Triage
Duplicate, wrong repo, too thin to start, or ready.
02 · Sort
It applies the labels you already use. Sorting never starts work by itself.
03 · Supplement
Repro steps, related links, acceptance criteria.
It labels and comments. Your people close. Every pass runs on your model key.
Your save file
Inventory
$ 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 whatRules of the save
Every memory records its source. When a person and Marshall disagree, the person wins. Cancel and you keep the brain.
Quest log
Install the GitHub App. Create or connect a brain repo.
Anthropic or OpenAI, on your own account. No resale.
One runner, one tracker, one chat.
Confirmed answers land in the brain, with their source.
Your ready label starts the run.
Map
Coming soon: RepoStaff will host the Management UI. Your agents will run on compute you choose.
Coming soon. Old machines in a closet count. Each task gets a fresh VM.
Coming soon. One CloudFormation stack you review and create. Cost estimate on the pricing page.
Coming soon. After self-run. Waitlisted, no date.
Tap what your team uses. Choose one per group to check fit.
Cloud compute is planned for customer AWS accounts after Linux/KVM hardware. Azure and Google Cloud are not supported. All integrations are pre-launch.
Your stack is coming soon.
GitHub, Linear, Slack, AWS. These integrations are planned, not operational.
Get startedYour Anthropic or OpenAI API key. No subscription logins. No token resale or markup.
Shop
Planned prices; they may change. Model usage on your own key and your own compute are billed separately. Full pricing
$499
RepoStaff's AWS. 100 VM-hours included, then metered up to your spend limit.
Hosted compute waitlist details$49
Up to three members on their own hardware.
Self-hoster rate detailsGood fit: 5–100 engineers · Code on GitHub · Already using AI coding tools · US-based
Not supported yet: Jira · GitLab or Azure DevOps · Regulated enterprise · Data residency guarantees
Code location, model keys, SOC 2, Devin and more: read the FAQ.