2buildDocumentationGitHub
2build / DocumentationRead Markdown (.md) View source

Product reference

English | Tiếng Việt | 中文 | 日本語 | 한국어

2build — product engineering, from a plan to a verified handoff. Concept illustration.

<img src="assets/2build-mascot-transparent.png" alt="2build mascot: an orange builder with a hardhat and a checked building block." width="128" height="128" align="right">

Review the prototype. Let Autopilot do the rest.

Start with Autopilot · Read the docs · Blog

2build helps your coding agent finish a feature. Give it a requirement, review the UI prototype, then let Autopilot write code, check it and fix issues.

For a small UI task, aim to see a prototype in ~5 minutes; timing depends on the task, model and project.

  • Quality: code is reviewed, tested and fixed before handoff.
  • Productivity: the agent handles each step without constant reminders.
  • Efficiency: adjust the prototype early to reduce rework.

Works with Claude Code, Codex, Antigravity, OMP and Grok, and any coding agent that supports Claude Code skills.

A 2found product. Previously named babysit; the bbs CLI, bbs: skills and existing .babysit state remain compatible. See branding and naming.

</a>

Install with one prompt

Paste this into your coding agent with terminal access:

text
Install 2build for the coding agent I am using. Follow
https://raw.githubusercontent.com/2found/2build/main/docs/install.md. Detect my OS and
current agent, reuse a working bbs installation or install the CLI, then install the
skill pack for this agent only. Verify bbs --version and the installed skills; report
any missing prerequisite instead of claiming success. Tell me whether I need to restart
and give me the exact Autopilot invocation for my agent to run my first small task.

You need a supported coding agent and its existing model access. Claude Code and Codex also need their CLI on PATH. 2build has no separate model account to configure; your agent's normal usage charges apply. Orca is only required for Team. See installation and troubleshooting.

<details> <summary>Prefer terminal commands? Homebrew on macOS or Linux</summary>
bash
brew tap 2found/2build https://github.com/2found/2build
brew install 2found/2build/bbs
bbs install

bbs install auto-detects Claude Code, Codex, Antigravity and OMP. For Grok and other compatible agents, follow the skill installation guide. Use bbs install claude, bbs install codex, bbs install antigravity, or bbs install omp to select one. Restart that agent after installation. Without Homebrew, use a release archive; on Windows, use WSL.

</details>

Start a new project

Nothing built yet? Open an empty folder in your agent and give it the idea (or none, and it asks):

text
/bbs:autopilot "A booking page for my yoga studio"

Autopilot writes the requirement, shows a clickable prototype for your one approval, scaffolds the services from the starters below, prepares the local setup and the deploy configuration, runs review and QA, and writes docs/BACKLOG.md: every feature as a ready /bbs:autopilot "…" line. It makes one skeleton commit and provisions and deploys nothing. Then paste the backlog lines one at a time to build each feature, and run /bbs:release when you are ready to go live; the first release is where accounts, a server and a domain are set up, with your yes.

2build starters provides versioned project templates with an agent harness, architecture guidance, tests and local QA. The first template is hono-bun: a Bun/Hono API with TypeScript and Zod.

The workflow above runs bbs bootstrap for each service. bbs bootstrap and bbs starter check are available in CLI 1.95.0+. The stable starter is ready to download directly from GitHub. Follow the starter quick start for commands and prerequisites.

</a>

Try your first ticket

  1. Restart your agent, open an existing Git repo and choose a small bug or feature with a clear success check. Autopilot works on your current checkout and leaves its changes uncommitted (a new project's skeleton commit is the one exception); create your preferred branch first if you want isolation.

  2. Invoke the skill in agent chat, replacing the example with your task:

    AgentExample
    Claude Code/bbs:autopilot "Add a saved-search screen. Show a prototype first, then implement, review and QA."
    Codex$bbs:autopilot "Add a saved-search screen. Show a prototype first, then implement, review and QA."
    OMP/autopilot "Add a saved-search screen. Show a prototype first, then implement, review and QA."
    Grok/bbs:autopilot "Add a saved-search screen. Show a prototype first, then implement, review and QA."
    Antigravity / other compatible agentsAsk it to use the installed autopilot skill for your task.
  3. Review the plan and, for UI work, open the prototype. Autopilot then asks whether to build: approve, or say what to change in the layout, flow or scope, and it revises and asks again. Once approved it builds in the same session. To stop at this checkpoint instead, include --stop-after=plan in your first request, then follow the resume instruction in its handoff. To run unattended, add --auto (for example /bbs:autopilot --auto "…"): Autopilot reviews its own plan, asks nothing, sets its own goal when the agent lets it, and otherwise keeps working in the same session until the ticket is done or blocked.

  4. Let Autopilot implement, review, fix, test and run QA. When the ticket passes its completion gates it leaves the verified change uncommitted in your checkout, with evidence, for you to commit or hand to create-pr; a blocked run names the missing input. After a restart, give it the ticket ID from the handoff to resume.

No Orca, new worker session or project configuration is needed for this first ticket. You choose the session's model before starting. Inspect progress and recover a session.

Keep building with 2found

When your product needs more, choose the tool for the next job:

Your next stepExplore
Give recurring work to a small team of agentsSoot, powered by 2agent — Soots answer visitors, watch sources, write and follow up on schedule, and ask you only when a person is needed.
Deploy what you built2server — deploy and operate apps on infrastructure you own.
Help customers discover your product2market — a marketing workspace in development, bringing product context, content and channels into one flow.

Start with 2build for the engineering work; explore these products when the need comes up.

Pick the work

/bbs:autopilot carries one piece of work — a bug fix, a feature, or a long change that takes all day — through plan, review and QA in one session. Start there: one long session keeps its context warm and spends the fewest tokens. Hand a whole project to /bbs:team only when you want what a single session cannot give (below). For a focused job, invoke a skill directly.

You type a handful of commands, one per step: setup-project, design-ui, autopilot or team, release, and marketing-video. Autopilot and Team run the other skills for you. The skill index lists them by step.

</a> </a>

</a>

Team: a product team for a whole project

Team works like a product team you hire. You meet it twice:

  1. Brief — you give the requirement, confirm the spec it writes back (one batch of questions, then none), and approve its basic design: architecture, a prototype of the whole product, and the release phases.
  2. Handover — you review the finished product with its evidence, then give the release go or send feedback. In-scope feedback is fixed and handed over again without another meeting; a request beyond the spec comes back as a design change.

Between those, the team designs, builds, reviews, QAs and integrates on its own, one commit per phase on your local base, and asks you nothing. Workers never commit, and Team needs Orca. /bbs:foreman still works as an alias.

It costs more than Autopilot — a lead session, a design session and an integration session per phase — so use it for a reason:

  • You want to step away — two touchpoints instead of steering a session, and a reviewed, releasable product at the end.
  • Phased releases — production increments you allow at design time, instead of one large change at the end.
  • Independent verification — a fresh session reviews and QAs each phase; Autopilot reviews in the session that wrote the code.
  • Parallel work on disjoint parts — separate services or repos joined by a frozen contract, when wall time matters.

Without one of these, a single Autopilot run, even a day-long one, is faster and cheaper.

<details> <summary>CLI and configuration details</summary>

Skills and CLI

Skills are the agent-facing workflows; bbs is the companion CLI they use. /bbs:team runs the product team; bbs team (alias bbs foreman) manages model policy lookups, durable lead records, phases (integrate, land, release), suggested next steps, and reports; bbs contract manages the interface contracts parallel tickets build against. /bbs:autopilot runs the one-ticket workflow; bbs autopilot exposes its checkpoint and state helpers. bbs ticket owns ticket identity, DAG relations, evidence, test-surface leases, and adopting finished tickets' work.

Useful CLI entry points:

bash
bbs dashboard
bbs team report <parent-ticket>
bbs team next <parent-ticket>
bbs ticket dag <parent-ticket>
bbs autopilot snapshot --json
bbs autopilot recover --json

snapshot reads canonical ticket state and gate evidence; recover adds bounded artifact excerpts for resuming work.

Run bbs <subcommand> --help for usage. More detail: companion CLI, git flow, and operations.

Semantic decisions

semantic-decision makes task sizing, complexity, routing, test impact and review judgments an explicit reusable skill step. Its default provider is the current LLM; no configuration or extra model process is needed. To use Cloudflare:

bash
bbs config set semantic_decision_provider cloudflare
bbs config set semantic_decision_model clef-flash  # or clef

Set CLOUDFLARE_ACCOUNT_ID and CLOUDFLARE_API_TOKEN in the environment or ~/.babysit/.env. Only user configuration selects the provider; project policy can restrict external inference. Cloudflare errors or uncertain answers fall back to the current LLM. See the skill and contract. See settings without a dashboard control for config paths, defaults, project limits and human-review behavior.

</details>

Documentation

Install and troubleshoot · Choose a skill · Inspect progress · Coordinate a project · Check performance · Docs and blog on 2found.dev

For coding agents, llms.txt links directly to the Markdown guides and runtime contracts.

Repository layout

  • bbs — the multicall binary (gitignored build output; go build -o bbs ./cmd/bbs). Hooks are compiled subcommands, bbs hooks <name>.
  • .claude/skills/ — agent skills and workflow references.
  • internal/ — Go CLI and services.
  • web/ — dashboard SPA, embedded in release builds.
  • tests/, docs/ — verification suites and user documentation.

Telemetry

Skill usage is recorded locally as JSONL under ~/.babysit/analytics/; telemetry is the primary feedback channel for unattended runs.

License

MIT — see LICENSE.