Appearance
Skip to content
Open source · MIT licensed
Agents work the tickets. You keep the workflow.
Your agent, your machine, your subscriptions. bakers.build moves every ticket through the states your team already follows, with you deciding at every step.
Nothing to host to get started. Add worker machines when your laptop becomes the limit.
a ticket, moving through its states
AGENT
qualification
The agent reads the ticket, classifies it as bug or feature, and publishes a versioned handoff before any work starts.
AGENT
spec
Design decisions, an implementation plan and checkable acceptance criteria. You approve one scope before anything is built.
AGENT
coding
Implemented in an isolated git worktree on its own branch; the acceptance checks run before the agent hands the ticket over.
AGENT + YOU
review
A second agent gives an independent verdict with findings; you confirm it, or send the ticket back with what you expected.
YOU
done ✓
You authorize the merge. The ticket closes and its status is mirrored back to the customer's origin issue.
⌁
Your agent
Claude Code, Codex, Grok, Kimi, OpenCode or Pi, on your own subscription.
▣
Your terminal
herdr keeps a tab per ticket today; tuios is coming. Swap it, keep the tickets.
⎇
Your stack
Drupal with DDEV and docs ship with step skills; write one for any other stack.
§
Your rules
Checks, standards and methods live in WORKFLOW.md in your repo.
THREE WAYS IN
Start where you are
Developers start on their own machine, project managers in the browser, leads with one question: does it fit us?
With your agent in the repo
- 1Describe a ticket; your agent writes it with checkable criteria.
- 2Your conductor runs it on your machine, step by step.
- 3You read each summary and give your verdict in the browser.
In the browser only
- 1Connect claude.ai, Claude Desktop or Cowork once.
- 2Create and qualify tickets in your chat, up to the spec.
- 3Follow, answer and accept in the bakers.build web UI. Nothing to install.
For your team
- 1Connect the project once; your agent writes WORKFLOW.md.
- 2Encode your checks, standards and method; swap agents and tools freely.
- 3Add worker machines when a laptop becomes the limit.
OWNERSHIP
Every state has an owner
A ticket moves through explicit states. Agents qualify, spec, code and review; you approve the plan, confirm the result and authorize the merge. Ownership is your call, per state, in WORKFLOW.md.
CONDUCTORS
Runs where your code runs
A conductor is a daemon on any machine you own. It registers itself, claims tickets it can work, creates a git worktree per ticket, and drives the agent to the next state.
Watch a ticket get worked
A real run, condensed: coding, tests, review, a change request, and the merge.
BB-142 · Fix login redirectcoding
✓ conductor studio-m1 claimed BB-142 · worktree feat/bb-142
AGENT · CLAUDE Run intake: loading @bakers-build/implement-ticket, reading the spec handoff and the acceptance criteria.
AGENT · CLAUDE Implementation complete in the worktree. Running the acceptance validation before handoff.
✓ playwright · 12 passed, 0 failed · login-redirect.feature
AGENT · CLAUDE Verification summary published to the ticket. Writing the next state.
✓ state written: review · run retired by the server
HUMAN · REVIEWER The redirect works, but keep the return URL after login. Sending it back.
↺ state written: coding · changes requested
AGENT · CLAUDE Preserving the return URL now, regression scenario added to the feature file.
✓ playwright · 13 passed, 0 failed · keeps return URL after login
✓ state written: review
HUMAN · REVIEWER Review is OK. Merge it.
✓ merge authorized by you · merging !241
!241 · Fix login redirect
feat/bb-142 → develop
ARCHITECTURE
Four pieces, cleanly separated
A shared control plane, conductors on your machines, the agents you already use, and a workflow that lives in your repo.
Control plane
The shared source of truth for projects, tickets, and runs. Your conductors talk to it over an API; the cockpit UI reads from it live.
Conductor
A daemon on any machine you own. It registers itself, claims tickets it can work, creates a git worktree per ticket, and drives the agent.
Agents
The AI CLIs doing the work: Claude Code, Codex, Grok, Kimi, OpenCode, Pi. Each is an addon; swapping agents is a config line, not a migration.
Skills
The workflow itself, written as versioned skill files in the repo. What "qualify" or "review" means is code-reviewed, not tribal knowledge.
Claude CodeCodexKimiOpenCodePi
FAQ
Questions, answered
Which agents does bakers.build support?
Claude Code, Codex, Grok, Kimi, OpenCode and Pi ship as addons today, each running on your own subscription. An agent is a plugin, so adding another CLI is an addon package, not a fork, and choosing the agent for a machine is one line in its machine config.
Do I need to host something?
Like GitLab, you choose. Use a hosted bakers.build server and get an account there, or run your own: one server, the control plane, that holds projects, tickets and runs. Either way conductors and agents run on your own machines and connect outward to it.
Does my code leave my machines?
No. Conductors work in local git worktrees on hardware you own, and agents talk to their model providers under your own keys. The server stores tickets, states, handoff comments and each run's footprint. A self-run server also keeps run transcripts, unless you turn that off per conductor; hosted bakers.build keeps handoffs and footprint only.
Can a human own a workflow state?
Any state can be owned by a human or an agent, and you change it per state in WORKFLOW.md. The shipped setup lets agents qualify, spec, code and review; you confirm the plan, the result and the verdict, and you authorize the merge.
How is this different from a coding agent in my editor?
An editor agent helps one person in one session. bakers.build runs the whole lifecycle across a team: tickets queue up centrally, any registered machine picks them up, every run is recorded with its transcript and footprint, and the process itself is versioned in the repo.
What does it cost?
bakers.build is MIT licensed and free. You pay only what you already pay: your own infrastructure and your agents' model usage. Every run shows its tokens, turns and tool calls on the ticket.
Our stack is not Drupal. Does bakers.build still fit?
Yes. bakers.build does not care about your stack; step skills do. Drupal with DDEV and documentation ship today, and a developer writes a step skill for another stack once. Until then, documentation tickets and anything without build checks already run.
What if nothing starts?
Check three things on the ticket page: it names your conductor and that conductor is online, no blocking ticket is still open, and the header shows no reason such as frozen or failing. The first-ticket guide walks you through it.
How do I answer the agent when it asks me something?
In its terminal tab on the machine that runs the ticket; your workspace manager opens one per ticket. A comment on the ticket page reaches the next step instead.
Start with one ticket
MIT licensed. Your agent, your machine, your rules.