Your AI has speed. We give it direction.
harness-boot is a multi-agent development harness for Claude Code. Where most AI tools add capability, we add focus.
A loose horse runs fast in every direction. A harnessed horse runs fast toward something.
You ──▶ ① Convert ──▶ ② Evolve ──▶ ③ Focus ──▶ ④ Collaborate ──▶ ⑤ Unify ──▶ Result
(the context) (the docs) (the rules) (the experts) (two commands)
| # | Strength | How it works | What you get |
|---|---|---|---|
| 1 | Convert | Plain-language ideas convert into an intermediate language — structured specs that every AI agent can act on directly | Same context for every agent — less guessing, sharper output |
| 2 | Evolve | Edit one place, the rest stays in sync; mismatches surface automatically; your manual tweaks survive | Design docs are always current — no manual upkeep |
| 3 | Focus | Each agent works inside its lane; completion criteria are enforced by the system, not by trust | AIs stay on the work that's theirs |
| 4 | Collaborate | Role-specialized agents follow set procedures; every decision and disagreement is recorded | Blind spots get covered, every step is traceable |
| 5 | Unify | Two slash commands. Talk to it in plain English; see the plan before anything runs | Almost nothing to memorize |
Plain language / plan.md / existing code
│
▼
┌──────────────────────────────────────────┐
│ spec.yaml (single source of truth) │
│ ├─ Ideas — vision · users │
│ └─ Rules — features · decisions │
└──────────┬─────────────────┬─────────────┘
│ │
Auto-derived Expert collaboration
│ │
▼ ▼
┌────────────────┐ ┌──────────────────────┐
│ domain.md │ │ orchestrator │
│ architecture │ │ ├─ Planning │
│ events.log │ │ ├─ Design │
│ chapters/ │ │ ├─ Implementation │
│ drift detector │ │ ├─ QA · Integration │
└────────────────┘ │ └─ Audit (read-only)│
│ + ceremonies │
└──────────────────────┘
│
▼
/harness-boot:work
(no args = dashboard · words = intent)
In Claude Code:
/plugin marketplace add qwerfunch/harness-boot
/plugin install harness-boot@harness-boot
cd my-new-projectPick the entry point that fits — both feed into the same harness:
# A. From a one-line idea
/harness-boot:init "a simple to-do app"
# B. From an existing planning doc (plan.md, design notes, a sketch)
/harness-boot:init plan.mdThen drive every feature through the lifecycle:
/harness-boot:workIf it takes more than 5 minutes, open an issue. We'll fix it.
Use this when you want to run from a local clone — for contributors, forks, or offline setups. The repo's .claude-plugin/marketplace.json makes any clone act as a self-hosted marketplace.
Note: harness-boot is not listed in the official Claude Code marketplace yet — the
qwerfunch/harness-bootform above resolves to this GitHub repo directly.
git clone https://github.com/qwerfunch/harness-boot.git
cd harness-bootThen in Claude Code (use the absolute path of the clone):
/plugin marketplace add /absolute/path/to/harness-boot
/plugin install harness-boot@harness-bootTo update later, git pull in the clone and run /plugin marketplace update harness-boot.
Append whatever you'd normally say after /harness-boot:work. Short keywords and full sentences both work.
| Intent | What you can say |
|---|---|
| Dashboard / what's next | (no args) · "what's on my plate?" · "where am I?" · "what should I do next?" |
| Run checks | "test it" · "run the gates" · "check what's there" · "make sure nothing's missing" |
| Mark as done | "it's done" · "ready to ship" · "I think we're good" · "let's wrap up the login feature" |
| Block / pause | "I'm stuck" · "pausing this" · "the upstream API isn't ready" · "calling it for the day" |
| Start / switch | "login feature" · "let's do signup next" · "F-3" · "tackle billing first" |
| Resume | "keep going" · "back to yesterday's work" · "do the recommended one" |
| Add evidence | "I checked" · "got reviewed" · "a friend tried it" · "QA passed" |
| Drop a feature | "skip this one" · "cancel" · "F-9 — never mind" |
| Look back | "what did I do last week?" · "history of F-3" · "why did we go with OAuth?" |
These are just samples — Claude reads context and handles other phrasings too. When it isn't sure, it shows you the plan and asks Y/n before running anything.
A typical happy path — start a feature, run the gates, mark it done.
You> let's start the login feature
Claude> Starting Login (F-3).
To finish: smoke run (gate_5) PASS + at least 1 evidence.
Next: run the gates.
You> run them
Claude> tests (gate_0) PASS — 19 unit tests
smoke run (gate_5) PASS — evidence auto-saved
type/lint/coverage (gate_1/2/3) SKIPPED — no tool detected
Iron Law satisfied (smoke run + evidence). Mark Login (F-3) done? [Y/n]
You> Y
Claude> Done. Retro saved: .harness/_workspace/retro/F-3.md
Up next: Signup (F-4). Say "switch" or "show what's left".
harness-boot
Active: "Login feature"
Progress: 3/6 gates passed · 2 evidence
Blockers: accessibility · Space-key behavior unclear
Pending: "Signup" · "Forgot password"
Next: (1) run the next gate (recommended)
| Project | Preview | Demo | Source | Description |
|---|---|---|---|---|
| cosmic-suika |  |
Play | GitHub | Space-themed merge game |
| Yours next | — | — | — | Add your harness-boot project here |
Built something? Open a PR adding your project to the table — copy an existing row as a template. Or open an issue with the image and a one-liner if a PR is overkill.
Send any image, GIF, or screenshot that shows the project — plus a one-liner and a link. We'll optimize and place it on merge. Full guide: docs/assets/README.md.
Does harness-boot actually produce better output than vanilla Claude Code — fewer tokens, more issues resolved? The framework for an apples-to-apples comparison is public, with the methodology, task list, and reproduction scripts open for inspection.
- Suite: SWE-bench Verified (20-task subset across 9 popular Python repos: django, sympy, scikit-learn, matplotlib, sphinx, pytest, requests, flask, pylint, astropy, pandas, xarray)
- Measures: resolve rate · token consumption · wall time · code quality signals (LOC, tests added, drift catches)
- Status: framework public, measurement in progress
- Read:
docs/benchmark/swe-bench-verified/REPORT.md— results table fills in row-by-row as runs complete - Methodology:
docs/benchmark/swe-bench-verified/README.md— why SWE-bench Verified, why 20 tasks, what's measured - Reproduce:
docs/benchmark/swe-bench-verified/scripts/setup.md— end-to-end external setup - Validity caveats:
docs/benchmark/swe-bench-verified/analysis/threats-to-validity.md - Privacy: the auto token capture (Stop hook · F-174) reads only the local transcript file Claude Code already wrote and appends token counts + the model id to your local
events.log. No network access. SetHARNESS_DISABLE_TOKEN_HOOK=1to opt out entirely.
harness-boot/
├── .claude-plugin/ plugin.json · marketplace.json
├── agents/ specialist agent definitions
├── commands/ slash commands (init · work)
├── hooks/ session-bootstrap · prompt-log
├── src/ TypeScript implementation
├── dist/cli/ esbuild single-file bundle (committed; no node_modules at install site)
├── bin/harness Node shim that loads the bundle — `harness <subcommand>`
├── self_check.sh 5-step self-dogfood verification
├── skills/spec-conversion/ plan.md → spec.yaml conversion
├── docs/ schema · templates · samples · portfolio assets
└── tests/parity/ TS parity test suite
v0.15.0 — Adaptive drive + structural archive separation + plugin-level writing rules (F-129 → F-140). Iron Law gains paired perf_regression / perf_resolved evidence kinds so silently-regressing perf cycles can no longer ship. complete() now relocates done feature bodies into a sibling spec.archive.yaml and harness sync migrates pre-existing bodies in one pass — spec.yaml stays slim by construction. harness drive becomes adaptive: deterministic replan after each completion, periodic real-test injection (harness.yaml drive.real_test.command), and transient-retry on flaky e2e. Plugin-level docs/communication-rules.md standardises answer-first format and native-tone replies for any user language.
Previously: v0.14.3 — Stability + cleanup batch (F-122 → F-128). tests/parity/driveLoopAndPlan.test.ts time-bomb fixed (F-122) so CI on every downstream fork stays green; CLAUDE.md brought back in sync (F-123) and refactored into stable / semi-static sigil / volatile-pointer layers so future drift is structurally prevented (F-124); the pre-cutover Python footprint (legacy/scripts/, tests/unit, tests/integration, tests/scale — ~28 k lines combined) evicted from git tracking, working trees preserved locally (F-125 + F-126); hooks/pre-commit-phase2.sh now short-circuits during git's own merge / cherry-pick / revert / rebase finalizers so conflict-resolution commits no longer require HARNESS_BYPASS_PRE_COMMIT=1 (F-127). Zero capability addition, zero behaviour change; v0.14.2 installs are functionally identical.
v0.14.0 — drive ships (F-118 + F-119). New autonomous-loop slash command /harness-boot:drive "<natural-language goal>" that scaffolds a Goal (a container that groups N features) via researcher → product-planner → feature-author and then runs every feature's gate cycle to completion. Bounded by design: BR-015 forbids self-issued --hotfix-reason, git commit/push/tag, and any shared-state mutation; the loop halts on 9 enumerated conditions (commit boundary, retry threshold, drift error, blocked feature, wall-clock, iteration cap, network failure, STOP file, plan-phase approval) and yields back to the user. Read-only harness drive --status [G-N] [--all] [--json] [--watch] is mtime-invariant (CQS, BR-012).
Previously: v0.13.2 — Repo root cleanup (F-117): dead Python config (pytest.ini, requirements-dev.txt) removed after the v0.13 TS-only cutover. No behavior change.
v0.13.1 added the feature-author skill — auto-triggers on Korean ("X 기능 구현해줘", "X 만들어줘") or English ("draft a feature", "spec out X") prompts to scaffold a complete features[] entry with shape detection, project-mode-aware AC count, and lockstep paste instructions.
- Changelog — CHANGELOG.md
- Developer guide — CLAUDE.md
- Issues — GitHub Issues
# For contributors building from source (devDependencies only —
# end users install via /plugin install and need no npm step):
npm install
npm test # vitest suite (parity tests)
bash self_check.sh # 5-step structural verificationMIT — Free to use, free to fork.