LEAPs

A local-first product-discovery tool. LEAPs takes any fuzzy idea and grinds it into a conviction-ready decision through a five-stage pipeline — Learn → Explore → Assess → Prove → Ship — scored by a rigor-gated conviction model (0–100, Pursue / Needs more / Kill) that can only run high when claims are backed by real, sourced evidence.

It runs on your machine, connects to your tools (BI, analytics, design, tickets — anything with an MCP server), and grounds every question in your company and role. No accounts, no hosted backend, no SSO — clone it and go.

Built for PMs, CPOs, and commercial folks doing light discovery to build conviction or kill an idea — instead of debating PRDs and decks.

Requirements

• Claude Code (the claude CLI) — installed and signed in (run claude once, then /login). LEAPs drives your own local Claude to run discovery, so this is required for the conversation pane to pull evidence and move conviction. (Without it you can still browse the board and the sample concept — you just can't run live discovery.)
• Node.js 18+ — 20 is recommended (pinned in .nvmrc).
• macOS, Windows, or Linux to run from source. The packaged build (npm run build) targets macOS on Apple Silicon.

Quick start

git clone https://github.com/tonyfadel23/leaps.git
cd leaps
npm install
npm start          # launches the desktop app (Electron)

On first run, a short setup walks you through signing into Claude, pointing LEAPs at your concepts folder and tools, an optional design-system connect, and your business context (company, role, goals) — then drops you on the board with a sample concept to explore. (Conviction thresholds and everything else are editable later in Settings.)

To build a packaged macOS app (Apple Silicon):

npm run build      # -> dist/mac-arm64/LEAPs.app  (unsigned — see "Staying up to date")

How it works

• Concepts are folders. Each idea lives in pipeline/{slug}/ as plain files (markdown, JSON, HTML prototypes). Nothing is locked in a database.
• The pipeline is Claude Code skills. /learn, /explore, /assess, /prove, /ship are conversation-first skills under .claude/.
• Conviction is one engine. electron/engine/conviction.js scores Problem, Size, Feasibility, and Evidence quality into a single verdict. The brief shows it; /eval audits it (it cannot be talked up, only sourced up).
• Live discovery drives your own Claude. The conversation pane spawns your local claude CLI inside a concept folder, using your auth and your connected tools to actually pull evidence and move conviction.

Connect your tools

LEAPs wires data sources into discovery via MCP servers declared in connectors.yaml (a sample ships as connectors.example.yaml). Map a semantic role (e.g. metrics_source, knowledge_base) to any MCP — Looker, BigQuery, a data gateway, Figma, Jira, Slack, etc.

# in the app's conversation, or a terminal Claude session:
/setup             # scan your MCPs and write connectors.yaml

Settings shows which sources are wired. Your filled-in connectors.yaml is gitignored — keep internal endpoints out of commits.

Give it your context

Copy .claude/skills/_shared/reference/business-context.example.md to business-context.md and fill in your company, markets, OKRs, and your role & scope. The skills and the discovery agent read it to ground every question in your business — and the pm-* agents adopt the decision lens you describe. (Your filled-in copy is gitignored.)

Staying up to date

• Running from source (recommended): git pull && npm install. That's the whole update.
• Packaged app: because it ships unsigned, macOS blocks it on first open — right-click the app → Open (or System Settings → Privacy & Security → Open Anyway) once, and it runs normally after. LEAPs checks GitHub Releases on launch and via Check for Updates… in the menu, and points you to the new build. It does not silently self-install — that needs a code-signed app, and LEAPs ships unsigned (free OSS). If you fork and sign it, swap electron/engine/update.js for electron-updater against the same releases feed; the seam (check → notify) is unchanged.

Testing

npm test          # engine/script unit tests + structural skill/agent checks
npm run eval      # + contract tests over your pipeline ideas

Skills and agents are prompts, so they're validated by a three-tier pyramid (structural → contract → LLM-as-Judge via /eval) instead of unit tests — see evals/agent-tests/README.md.

Roadmap

LEAPs is local-first today. Multi-person collaboration on a concept (Google-Docs / Miro style) is the next milestone — see docs/collaboration.md for the design.

Contributing

See CONTRIBUTING.md. Issues and PRs welcome.

License

MIT.