Knit gives your coding agent a whole team's discipline (product thinking, a lean plan, design polish, tests,
review, docs and a live deploy) at the cost of one normal session. It works in Claude Code, Codex, Cursor,
OpenCode and Antigravity. One round of clickable questions, a one-page plan you approve, then the build, slice by
slice. Every step is logged, and nothing goes live without your OK.
Same cost as a plain session: on the same calculator, a plain session used 117,960 tokens and Knit used
124,120 (+5%), with less code, more checks, a plan, a run log and a README. Benchmark.
How it works
- Intake. One round of clickable questions: must-haves, stack, where it goes live, solo or team.
- Plan. One page: stories with acceptance checks, the look, the build, the tasks. The Lean checklist cuts what doesn't need to exist. You approve the plan.
- Build, slice by slice. Your session puts on each role's hat (Designer, Builder, Tester) using short checklists, with tests only where they pay.
- Check and ship. Tests, lint and build, one Reviewer pass (bugs, security, lean, accessibility, polish), a README with the badge, and a live deploy. You approve going live.
Solo or team. Solo is the default: one session wears every hat, so it costs what a normal session costs. Team mode is opt-in: the specialist agents below work in parallel for speed, and Knit tells you up front that it costs more, because every agent re-reads the project.
The team (each role is a checklist in solo mode, an agent in team mode)
| Agent | File | Model | Job |
|---|
| Manager | skills/knit-build/SKILL.md | your session | Intake, build mode, gates, delegation, loops, final validation |
| Product Manager | product-manager.md | opus | PRD, US- user stories, acceptance criteria |
| Architect | architect.md | opus | Architecture, openapi.yaml contract, tasks grouped into chunks |
| Lean Coder | lean-coder.md | sonnet | Plan check and trim review: the best code is the code nobody had to write |
| UI/UX Designer | ux-designer.md | sonnet | Flows, screens, design system and tokens, visual check |
| Frontend Developer | frontend-dev.md | sonnet | Web UI from the design and API contract |
| Backend Developer | backend-dev.md | sonnet | API, auth, data access, migrations |
| Tester | tester.md | sonnet | Unit, API and Playwright e2e tests mapped to stories |
| Debugger | debugger.md | opus | Root cause of failures before anyone fixes |
|
The Manager is a skill run by your main session, because subagents can't start other subagents.
Change any agent's model: line to suit your plan.
Build mode: in chunks or all at once
| In chunks (default) | Whole project |
|---|
| Best for | Anything bigger than a small app | Small, well-defined apps |
| How | Chunk 0 foundation, then thin end-to-end slices of 1 to 4 stories | One plan, one build |
| You see working software | After chunk 1, then after every chunk | At the end |
| If something is wrong | Caught early, small fix | Caught late, bigger fix |
A team-mode run
Team mode's first test built a local habit tracker with 11 agents. It worked, but it cost 2.16M tokens, which is
why solo mode became the default. Every number below comes from its run log.
| 11 agents · 22 runs | 2h 12m of agent time | 2.16M tokens | 197 unit/API tests | 112 / 112 e2e | Review: approved in round 2 |
|---|
flowchart LR
PM[Product Manager] -->|prd.md| AR[Architect]
AR -->|openapi.yaml| SE[Security Reviewer]
SE -->|threat-model.md| AR
AR -->|architecture.md| UX[UI/UX Designer]
AR -->|tasks.md| OP[DevOps]
UX -->|design.md| FE[Frontend Dev]
TE[Tester] -->|test-report.md| DB[Debugger]
DB -->|debug.md| FE
CR[Code Reviewer] -->|review.md| FE
Read the full run report and token report.
This is the beginning of the README. Read the rest at the original. Read the full README ↗
