For customers who want to see under the hood. You don't need any of this to use ShipForge — but if you're curious, here's the architecture.
Your app comes in. An evidence-backed launch plan comes out. Deterministic scanners and AI coordinate the analysis; automated quality gates check the findings before they reach you.
ShipForge uses a model-routing layer that selects approved providers based on quality, cost, latency, modality, and availability. New providers are added when they win against existing ones — customers don't feel any of it.
We benchmark every candidate model against our own tasks — code, audit, planning, creative. Only winners are promoted.
Routine work runs on cheaper models. Serious work uses the top of the class. You don't pay for what your task doesn't need.
Interactive tasks favor faster models. Deep audits favor slower, more capable ones.
Images, video, vision and long-context work each route to specialists.
If a provider degrades, ShipForge falls back to another. No dead ends.
This list changes. New models get promoted whenever they win a benchmark. We publish the current lineup so you know what's approved to run against your project.
Primary Brain / Orchestrator
CTO / team leader
Reserved for serious work. Not every cheap task should hit the strongest reasoning model.
Fast, Multimodal, Cost-Efficient Worker
Fast analyst
The router picks the best cost/performance Gemini variant per task — never a hard-coded model.
Code Specialist
Senior coding engineer
Purpose-built for repository-level coding, tool use, and long-horizon agent workflows.
Engineering Challenger / Backup
Second senior engineer
Deliberately not used on every task. Engaged when a second opinion or a fallback is genuinely warranted.
Creative / Image Specialist
Visual designer
Fast, photorealistic, strong instruction adherence — tuned for the Launch Studio asset workflow.
Video Specialist
Motion designer
Text-to-video and image-to-video, 720p/1080p, 2–15 s clips, synchronized audio, multi-shot generation. One video engine at launch — Veo can be added later only if it wins a benchmark.
We don't rank providers on this page ("A beats B") because model performance changes constantly. What matters is that ShipForge always routes to the current best available option for the task.
For the things that are objectively right or wrong — exposed secrets, missing account deletion, disabled Row Level Security — ShipForge runs real code, not an AI opinion.
For high-stakes work — security, database changes, payments, App Store compliance, production architecture — ShipForge is designed to run an additional cross-check before the answer reaches you. This automated judge is a planned capability; during Founding Beta, high-stakes findings are checked by an independent model challenge and automated quality gates instead.
Deterministic Scanner
Finds facts (missing RLS, exposed keys, absent deletion flow, etc.)
Primary AI
OpenAI Architect reasons about impact and drafts the fix.
Second Provider (when warranted)
GLM-5 provides an independent engineering opinion.
Quality Judge
Cross-checks scanner facts against the AI proposal before it reaches the user.
Provider requests are made through ShipForge — with your project scoped, your data minimized, and no training on your code. Restricting routing to specific providers is a planned capability.
Your code is never used to train any model. Provider requests are made with training disabled where the provider supports it.
Restricting routing to a chosen subset of providers — for example, US-only or EU-only — is a planned capability.
Your keys never touch the client. ShipForge's own provider keys are managed server-side and are never exposed.
You'll never touch any of this. It just works.
Run Free Launch Score