distilr

Terminal agent that distills any SaaS into a complete build spec — research, design tokens, architecture, and a phased plan, ready for your AI coding agent.

View on GitHub Quickstart

what it does

Point it at any SaaS — a form builder, a scheduling tool, a project tracker, an email tool, anything with a focused feature surface. distilr explores the public marketing site and the authenticated app in a real Chromium browser, distills what it sees into a feature catalog, walks you through a wizard, and emits a phased build spec your coding agent can execute.

You stay in the loop. The agent never enters payment info, sends emails, or clicks destructive buttons without your explicit terminal approval. Press p at any point to interrupt and steer.

scope. distilr produces an MVP-sized inspired-by spec — typically 5–10 phases, ≤ 8 must-have features. It's not a tool for reproducing sprawling enterprise platforms or creative-app suites feature-for-feature. Narrow your target to a specific module — e.g. just the support-ticket surface of a CRM, or the color-correction surface of an image editor.

how distilr respects the source

distilr studies a SaaS to inform a new MVP — never to clone it. Every agent in the pipeline operates under these rules, baked into their system prompts. The source of truth is one file (src/agents/guidelines.ts) so the prompts and these docs cannot drift apart.

  1. Never copy source code (even partially). We capture behavior and shape; the implementing agent writes new code from scratch.
  2. Never reproduce exact UI designs, layouts, or assets. Describe patterns, not pixel-perfect copies. No extracting icons, illustrations, custom fonts, or brand assets.
  3. Never copy text content — docs, onboarding flows, microcopy, error messages, marketing copy. Describe what the text accomplishes; new copy is written for the new product.
  4. Never use the source's trademarks or branding in a confusing way. The output is inspired-by, not a clone or knockoff. Pick a name that's clearly distinct.
  5. Respect the source's Terms of Service. Stay on the user-facing surface. Respect rate limits. Don't reverse-engineer protected APIs. Don't bulk-scrape.

These are not opt-in — they're injected into every heavy agent's system prompt at build time.

built for the agent era

multi-provider

One CLI, four providers. Anthropic, OpenAI, Google, OpenRouter — pick whichever you have keys for.

tool-agnostic output

Emits the harness-engineering doc layout your coding agent already knows. AGENTS.md and CLAUDE.md in one shot.

scope-aware

Pre-flight LLM realism check, hard caps on must-haves and phases, token-budget warnings at 50k / 200k / 500k.

you sign in

Real headed Chromium browser. The agent never sees credentials. Destructive clicks always prompt the terminal first.

checkpointed

Ctrl-C is always safe. Resume any project from its last completed stage.

open source

MIT-licensed TypeScript. PRs welcome — see CONTRIBUTING.

the pipeline

~/distilr — pipeline
1. target pick the SaaS to study (or describe your idea) 2. recon agent explores the public marketing surface 3. login headed browser opens — you sign in manually 4. explore agent drives the authenticated app 5. synthesize observations → structured feature catalog 6. wizard 10 questions to shape your MVP 7. architect system design + tokens + phased plan 8. emit harness-engineering doc tree, ready to ship 9. build setup checklist + copy-paste prompts for your AI coder

works with any coding agent

Output is tool-agnostic — same content under both AGENTS.md and CLAUDE.md, plus per-phase exec plans your agent can pick up.

quickstart

~ — bash
$ git clone https://github.com/ghoshsanjoy78/distilr.git
$ cd distilr
$ npm install
# downloads Chromium + compiles TS via postinstall
$ ./distilr

Requires Node 20+. The first run launches a setup wizard for your provider and API key.