SpecLeap
Framework

No. SpecLeap is provider-agnostic by construction, not as a later add-on. The framework makes no calls to any model API: zero traffic to api.anthropic.com, api.openai.com or any other provider.

All the project's "knowledge" lives in .md and .sh files inside your repo. Each assistant reads its own instruction file: Claude reads CLAUDE.md, Cursor reads .cursorrules, GitHub Copilot reads .github/copilot-instructions.md, and so on across 9+ supported assistants.

The specification is the asset, not the model that processes it.

Vibe coding is fast at the start and expensive at the end. After 50-100 iterations the AI loses context, contradicts existing patterns, reinvents utilities. You become the long-term memory system.

SDD captures the project's intent once, in an immutable CONTRACT. The AI can iterate 1000 times but always returns to the same north. Vibe coding scales with developer time; SDD scales with spec quality.

There's a minimum where SpecLeap pays off: roughly one week of real work.

• → Weekend side-project: probably not — the initial survey is overhead you won't recover.
• → 1-4 week project: yes, clearly — a single "no, this isn't what I meant" cycle already costs more than the 20-min survey.
• → Multi-month with several contributors: fundamental.
• → Enterprise / regulated: especially — versioned spec traceability is what audits ask for.

Yes, it works with Cursor, Continue.dev, GitHub Copilot, Gemini, Codeium, Tabnine, Cline and any assistant that reads project instruction files.

What you lose without Claude Code: native integration with the 34 Agent Skills (in other assistants the .md files remain accessible but don't auto-load) and the spec-guard.sh hook that ties into the Claude Code runtime (the equivalent protection is covered by the independent pre-push git hook).

What you DON'T lose: CONTRACT, agents, commands, memory-bank, specs, Asana/CodeRabbit integration, the survey, the openspec CLI — all the same.

You do it without losing anything. But the "how" depends on your setup, because AI provider, assistant and IDE are three distinct layers:

If you use Cursor or Continue.dev, switching provider is a click in their settings. If you use Claude Code and want to move to OpenAI, you swap extensions (not necessarily IDEs). If you switch IDE entirely, you open the repo in the new IDE and the full context is still there. The .md and .sh files stay intact no matter what.

For some commands yes (crear-tickets, planificar, status updates inside implementar). For the bulk of the framework, no.

Without Asana these still work: survey, CONTRACT, agents, commands refinar / documentar / explicar / compliance / adoptar, the 34 Agent Skills, the full openspec CLI, memory-bank, global standards and the 3-layer validation pipeline. If your team uses Linear, Jira, Notion or GitHub Issues, you can manage the backlog manually.

The CONTRACT is immutable by design, so classic conflicts on that file don't exist. Changes to the original scope go to ANEXOS.md (append-only, versioned). Technical decisions go to context/decisions.md with date and author.

Each feature in its branch feature/<project>-<ticket>-<desc>, PR against stage, then PR against main. CodeRabbit reviews against CONTRACT + standards. The 3-layer pipeline validates before merge.

No, it complements them. Each operates at a different layer:

• → SDD (SpecLeap): project's long-term memory — what's being built and why.
• → TDD: how you write code (test first).
• → Agile / Scrum: how you organize work over time.

SpecLeap is the missing layer in the methodology stack now that AI writes the code. They coexist without friction: your CONTRACT doesn't clash with your sprint board, your TDD tests don't clash with your specs, and your Scrum Definition of Done is enriched by SpecLeap's 3-layer validation pipeline.

SpecLeap is free and open source (MIT). No account, no paid plan, no card. Clone the repo or install it with npm and the code is yours.

The only cost is the assistant you already use — Claude, Cursor, Copilot or another. SpecLeap does not sit in the middle of that subscription or add one of its own.

You need Node.js 18 or newer, Git and Bash. Full setup takes 10-15 minutes. PHP, Composer, PostgreSQL or Docker only matter if the project you are about to build uses them.

On Windows, install Git for Windows: it includes Git Bash and the installer finds it on its own. If Windows has the folder locked while renaming, setup continues with the current name. The reliable path is to clone with the final name: git clone <url> my-project. macOS and Linux already ship Bash.

Yes. The adoptar command reads the code you already have and writes a retroactive CONTRACT: what exists, what is out of scope, and which rules apply from now on.

It does not rewrite the application. It documents what is already built so later changes stop drifting in silence.

The survey saves progress every 10 questions. When you finish you can expand from MVP to Serious Project — adding governance, SLAs and compliance without repeating what you already said — or attach a free paragraph in Creative mode.

Once the CONTRACT is approved, it is not edited. A scope change is recorded in ANEXOS.md with the refinar command. That keeps the original ask separate from what was added later.

No. The standards and part of the skills are tuned for Laravel, React, TypeScript, PostgreSQL and shadcn, and the survey lets you declare a different stack. From then on the CONTRACT is what counts.

Quality, process, security and design skills still apply. The Laravel or React skills help less if your project does not use that stack.

It has no server and no account. The CONTRACT and the code stay in your repository. The framework does not call model APIs. During setup, the installer clones the 34 skills from public GitHub repos into ~/.claude/skills/.

What leaves your machine afterwards is what you connect: the AI assistant you already pay for, GitHub if you push, Asana if you paste the token, CodeRabbit if you install the app. Running the installer again only updates the catalog skills. A skill of yours, under another name, is left alone.