An AI team that keeps working when you don't.
Give Dobby a goal or a ticket. It plans the work, builds it, tests it and reviews it, then opens a pull request, on your own computer and your own AI plan. It paces itself to the limit you set.
Needs macOS or Linux (WSL2 on Windows), Node 24 or newer, git, the GitHub CLI, and Claude Code on a Pro or Max plan or an API key. Stop it with dobby pause, remove it with dobby uninstall.
On pace to use 72% of this week. 2 things need you.
14 merged in the last 7 days · 3 ready · resets Thu
Needs you
Working 3
Eight roles, each with one job.
Each role is a separate run with its own instructions. Here is one ticket passing through them.
#95 CSV export
Product Manager
Wrote the spec: which orders, which columns, who can export.
Engineering Manager
Queued it for today as a medium ticket.
Designer
Specced the Export button and its dialog.
Builder
Built it on its own branch. 18 tests pass.
Tester
Checked it against the done-when line. Passed.
Code Reviewer
Found that big exports time out. Sent it back once, then passed it.
Chief of Staff
Put “merged” in your evening brief.
Talent Scout
Proposed a rule for the Builder: stream big files.
Plus supporting roles that run when a ticket needs them, such as the Program Manager, the Onboarder and Brainstorm. Tap a role to stop the ticket there.
It leaves part of your Claude plan for you.
Dobby reads your usage every 10 minutes. It spreads the week across the 5-hour windows, pauses when a window reaches its share, and stops starting work at your weekly cap, 95% by default.
- 08:05Morning brief arrives
- 09:10Builder starts “Add CSV export”
- 10:40Reviewer finds a missing check. Sent back
- 11:05Merged #89
- 11:405-hour window at the 60% cap. Pauses
- 13:00Window resets. Resumes
- 19:00Evening brief
$ dobby limits → Running · 5-hour 34% (cap 60%) · week 49% of 95%
On an API key it caps spend instead, $20 a day by default. Every pull request shows what it cost. Nothing runs while your computer sleeps.
Every change is checked before it reaches you.
A tester and a reviewer check every change. Problems go back to the Builder, not to you. Dobby merges only the exact commit that passed review.
Show stock on the product page #87
dobby/142-show-stock → main
- Builder
Pushed 1 commit. 24 tests passed.
- Tester
Matches the done-when line.
- Code Reviewer
1 finding: the page breaks if the inventory call fails.
- Builder
Pushed a fix.
- Code Reviewer
Review passed.
- Dobby
Merged the commit that passed review.
The process grows with the risk
It learns from its own mistakes.
Once a day the Talent Scout picks one role, looks at what went wrong and proposes a better rule or skill for it. The change arrives as a pull request. Merge it with one tap, and every later run of that role uses it.
Talent Scout
Pull request: a better rule for the Builder
In the last 7 days, the Code Reviewer flagged the same mistake 3 times: a missing type check.
- Push when the tests pass.+ Run the type check, then the tests, then push.
You decide what merges on its own.
Dobby sorts every change into a category. Pick a preset and watch what waits for you. Merges stay off on a fresh install until you turn them on.
These decide merges. A deploy after the merge still waits for your tap unless you set it to deploy on its own.
Builders work in a sandbox either way: no network, no credentials. On Linux it needs bubblewrap.
$ cat ~/.ssh/id_ed25519blocked by the sandbox$ curl https://example.comblocked: no network$ git pushblocked: Dobby pushes after the checks
Each role loads open-source skills you can read.
24 open-source skills ship with Dobby, each read in full and pinned to the version read. A role's file says which ones it loads.
roles/reviewer.md --- name: reviewer skills: find-bugs, security-review ---
find-bugsSentryApache-2.0- security-reviewSentryApache-2.0
roles/builder.md --- name: builder skills: verification-before-completion, diagnosing-bugs, test-driven-development ---
verification-before-completionJesse VincentMIT
diagnosing-bugsMatt PocockMIT
test-driven-developmentAddy OsmaniMIT
roles/checker.md --- name: checker skills: verification-before-completion ---
- verification-before-completionJesse VincentMIT
roles/product-manager.md --- name: product-manager skills: prioritization-frameworks, outcome-roadmap, create-prd, pre-mortem, competitor-profiling, review-management, android-aso ---
prioritization-frameworksPaweł HurynMIT- outcome-roadmapPaweł HurynMIT
- create-prdPaweł HurynMIT
- pre-mortemPaweł HurynMIT
competitor-profilingCorey HainesMIT
review-managementAppeekyMIT- android-asoAppeekyMIT
roles/lead.md --- name: lead skills: sprint-plan, writing-for-agents ---
- sprint-planPaweł HurynMIT
- writing-for-agentsMatt PocockMIT
roles/scout.md --- name: scout skills: retro, writing-for-agents, harness-engineering, evaluation ---
- retroMatt PocockMIT
- writing-for-agentsMatt PocockMIT
harness-engineeringMuratcan KoylanMIT- evaluationMuratcan KoylanMIT
Sources and licenses are in THIRD_PARTY.md. Listing them is not an endorsement.
The commands you will use most.
You can also use the dashboard on your phone, Discord buttons, or Claude through MCP.
$ dobby add shop "Add CSV export" "Owners can download orders as CSV" Mwrite a ticket$ dobby idea shop "people drop off at shipping"a seed idea for the Product Manager$ dobby goalsyour goals in rank order, from GitHub milestones$ dobby pause merges for 2hpause one part, for a while$ dobby hint 95 "reuse the report exporter"a hint for the ticket's next build$ dobby takeover 95pause the ticket and leave its branch to you$ dobby cost 95what the pull request cost, by stage$ dobby receiptwhat shipped this week on your plan$ dobby mcplet Claude read status and file ticketsPR comment @dobby reviewa review on any pull request, yours too