From raw idea to deployable codebase in one session.
SaaSForge discovers open-source tools, scores their SaaS potential on 5 weighted dimensions, assembles the best of them into a single product proposal, judges whether that product deserves to exist — GO / NO_GO / RISKY — then generates a ready-to-run project scaffold in under 5 minutes.
Both halves of starting a SaaS — judgment and boilerplate — automated.
The manual way
- Hours searching GitHub, judging maturity project by project.
- Research judgment that is inconsistent, undocumented — and lives in someone's head.
- The first week of every build spent re-creating the same skeleton: models, services, routers, migrations, workspace UI.
The SaaSForge way
- One pipeline: discover → evaluate → recipe → feasibility → scaffold. Each stage feeds the next.
- Every judgment quantified, auditable, and reproducible — with named competitors and prices.
- A hard veto layer a good pitch cannot argue past — margin under 20% is an automatic NO_GO.
Five stages. One verdict. A runnable codebase.
Every AI-driven stage is credit-gated, runs as a background job when slow, tracks partial failures explicitly, and refunds credits automatically when it fails outright.
Discover
Searches GitHub and HuggingFace Spaces in parallel — full metadata, licenses, freshness, community stats. Automatic categorization into 6 buckets. Star-velocity trending surfaces fast-rising repos that raw star counts bury; any tool jumping >50% in a week gets a spike flag.
1 creditEvaluate
Scores any tool on a 5-dimension SaaS-Readiness scale (total /100). Quick mode for cheap triage across a whole list; full mode returns recommendation, reasoning, strengths and weaknesses, a named competitor list with prices, and an MVP checklist.
3 quick · 15 fullRecipe
Given a target domain, assembles 3–5 open-source tools into one coherent product proposal: each tool gets a pipeline role, AGPL is excluded for clean commercialization, and the output carries pricing, cost per user, margin, and a 3-phase execution plan with week estimates.
20 creditsFeasibility
The veto layer — deliberately the last word. Interrogates the proposal: market saturation, unit economics (margin <20% → automatic NO_GO), build time, and worst-case bounded loss. Returns GO / NO_GO / RISKY with a confidence %, ranked risks and opportunities, and concrete final advice.
IncludedScaffold
The flagship: manifest-first design, then file-by-file generation with per-file failure isolation — one bad file never sabotages the project. Ships a downloadable ZIP plus ZCODE_TASK.md, a ready-to-paste developer handoff with priorities, do-not-touch lists, and measurable success criteria.
30 creditsSaaS Readiness, quantified out of 100
Five weighted dimensions. Paid competitors get named with prices; maintenance status and stale commits get flagged; licenses get scored for commercial viability.
Bar lengths show each dimension's weight, not a score. MIT / Apache = 10 · GPL = 5 · AGPL = 1 · none = 0.
Three production runs, transcripts included
| Run | Result |
|---|---|
| OpenManus — discovery + full evaluation | 73/100 · GO — 57,403★ MIT Python; 5 named competitors with pricing; $10–20/user/month cost model; 10-week MVP estimate; a 152-day-old last commit honestly flagged. Controlled contrast: without its README the same engine returned 22/100 · NO_GO — the scoring reads evidence, not vibes. |
| ContentFlow AI — recipe + feasibility | RISKY · 68% — 5 tools assembled with pipeline roles, compatibility 82/100, margin 89.2% — yet killed by competitor density 35/100 and a 15-week MVP competitors could outpace. The exact tension a human analyst would charge real money to surface. |
| KalamAI — scaffold generation | 19 files / ~6,400 lines — Arabic TTS product: 4 specialized backend services, a full Alembic migration, a 5-endpoint FastAPI router, 2 custom React components, README and ZCODE_TASK handoff — in under 5 minutes for ~$0.50 of AI cost, 95% per-file success. |
Metered like a utility. Refunded like a promise.
| Action | Cost |
|---|---|
| Basic search | 1 |
| Trending search | 2 |
| Quick evaluate | 3 |
| Full evaluate | 15 |
| Generate recipe | 20 |
| Generate scaffold | 30 |
| Watchlist · views · polling | 0 |
Where a hard, evidence-based veto already rules
Four high-stakes fields have converged on the same structure independently — a scoring layer separated from a veto layer that a persuasive memo cannot talk its way past:
VC & startup deal screening
A near-direct structural match. Investment committees already separate scoring (thesis fit, market, founder quality) from a decision layer with veto power — 17.1% of surveyed firms use explicit score-based voting where one strong enough 'no' kills a deal regardless of the memo. SaaSForge's margin-under-20% automatic NO_GO is that same mechanism in code. High-volume funds could screen deals before a partner's time gets spent; venture studios could reuse the full pipeline shape.
Real-estate development feasibility
The cleanest existing precedent for a hard numeric gate: debt-service-coverage thresholds are non-negotiable by lender type — roughly 1.15× for SBA, 1.25× for agency multifamily, 1.50× for life-company capital. A project that clears every other pillar but misses its capital source's DSCR floor does not get a 'everything else looked good' pass. That is SaaSForge's margin gate with a different name and number.
R&D stage-gate reviews
Almost exact vocabulary: Cooper's Stage-Gate — the dominant model in physical-product and pharma R&D — explicitly separates pass/fail 'must-meet criteria' from weighted 'should-meet criteria', producing Go / Kill / Hold / Recycle at each gate. That is SaaSForge's hard-gate-versus-weighted-score split, arrived at independently for the same reason: brilliant should-meet scores still die on a failed must-meet.
M&A target screening
Corporate-development teams already run structured scorecards across dozens or hundreds of targets, filtering hard on deal-breakers before scoring survivors on strategic fit. Same two-step logic as SaaSForge's discover-and-filter: drop AGPL candidates immediately, rank by metadata, reserve expensive full evaluation for what survives — cheap filters before costly analysis.
Honest failure handling, end to end
One AIClient for every AI call
3 attempts per call, 90-second timeouts, exponential backoff on 429/5xx, and a JSON-repair last resort. Every response carries an explicit partial flag: degraded results are saved and visibly flagged — never silently discarded, never faked success.
Async by contract
Slow work never blocks a request: POST deducts credits → creates a working row → dispatches the Celery task with a dispatch-failure refund guard → returns an id. The UI polls live progress — a real progress bar and file list, not a spinner.
Per-file failure isolation
In scaffold generation, a failed file becomes a clearly-marked error placeholder and generation continues. One bad file never sabotages the project; the failure is recorded on the row.
Owner-only, no existence leaks
Every evaluation, recipe and scaffold is owner-only; non-owners get 404, never 403. Credit mutations run under SELECT FOR UPDATE. All host ports bound to localhost; DB and Redis fully private behind a reverse proxy.
Model-agnostic gateway
All AI flows through OpenRouter — swap the model with one environment variable, zero code changes. External data from the GitHub Search API (5,000 req/hr) and HuggingFace Spaces API.
ZCODE_TASK — the handoff
Every scaffold ships a ready-to-paste developer task: what was built, what remains, ordered priorities, explicit do-not-touch files, and measurable success criteria ('docker compose up starts everything; login → input → output → download works end-to-end').
The bridge: opportunity → proposal → codebase
saas-template
The base every product is built on — sold on the Marketplace
idea-discovery
Mines 6 live sources for validated micro-SaaS opportunities
SaaSForge
This engine: the bridge from validated opportunity to working codebase
publisher
Turns a product URL into a full marketing content library
All five stages live in production
What's the hard number a good pitch shouldn't talk its way past?
Run your idea through the five stages — verdict in minutes, codebase in one session.