Forge
An AI engineering manager for every repo you have
A desktop app for Mac and Windows that reads every project in your coding folder, shows how far along and how production-ready each one is, and runs AI audits whose findings are checked against the code. Findings become plans you paste into your coding tool. It never writes to your repos.
getforge.dev
- Role
- Solo: product, desktop app, server, website, billing
- Platforms
- macOS & Windows app · Website & accounts
- Tech
- ElectronNext.js 16React 19TypeScriptSQLiteFastifyPostgreSQLBetter AuthAnthropic & OpenAI APIsPaddleDocker
122
commits since late September 2026
3
AI scans: structure, SEO, issues
32
Postgres tables behind accounts & credits
4
workspaces: app, server, website, shared
The problem
Many projects, one person, no engineering manager
Solo developers and small teams, especially those building with AI, end up with a folder full of projects and no clear picture of which are close to production, which have stalled, and what is quietly risky in each one.
AI code reviews help, but they can claim things that aren't in the code. Forge's rule is that every finding must point to the exact files and lines, and Forge checks that evidence itself before showing it.
01 · Scans
Measured first, then audited
Forge first measures each repo without AI: branch, uncommitted work, commits, CI, tests and deploy config. Then three colour-coded scans audit it: Structure, SEO and Issues. Each returns a report rated per topic, so you see at a glance what is good, fair or poor.
- A Facts tab with languages, packages, tests, CI and git activity, measured without AI
- An SEO score from 0 to 100, with the live site's pages checked too
- A risk overview for Issues, and the three things to do first
02 · Findings
Audits whose evidence is checked against the code
Three colour-coded scans (structure, SEO and issues) each return findings with a severity, a size and quoted evidence. Forge verifies every quote against the files: wrong line numbers are corrected, unsupported claims are sent back, and anything with nothing behind it is dropped.
- Findings sorted into Now, Next, Watch and Ignore lanes that survive re-scans
- “Check changes” re-audits only what changed since the last scan
- Evidence that can't be confirmed is marked, kept at low confidence and never put in Now or Next
03 · Plans
From finding to a brief for any coding tool
A finding turns into a Light, Standard or Deep implementation plan: objective, context, approach, steps with the files involved, and how to test it. Forge doesn't write code; you paste the brief into Claude Code, Cursor or whatever you use.
- Plans sized and priced before they run
- Copy as a brief, copy as Markdown, or download
04 · Manager
Ask about any project, typed or spoken
The manager answers questions across all your projects, cites the findings it bases its answer on, and offers the next action. Voice input is transcribed locally with Whisper.
- Context from analyses, findings and plans, not the whole repo
- Model chosen per job: Opus for audits, Sonnet for questions
05 · Privacy
The model only sees what it asks for
The AI never gets a repo as a whole. It asks for listings, searches and file reads, and Forge runs them locally with strict limits. Every task's files are listed in an activity log.
- Git-ignored files are invisible; .env files, private keys and certificates are never read
- Secrets pasted into ordinary files are replaced with [REDACTED] before anything is sent
- No API keys on your computer: AI is paid with credits through Forge's server
Under the hood
How it's built
- 01
Desktop app
Electron wraps a Next.js app and a custom Node server started on a free port with a fresh access token each launch; data lives in local SQLite. The account token is encrypted with the macOS Keychain.
- 02
AI gateway and credits
A Fastify server holds the provider keys, routes each job to the right Anthropic or OpenAI model, and charges a credit ledger in PostgreSQL. A pg-boss queue runs background work.
- 03
Evidence verification
Every quoted line in a finding is matched against the same redacted text the model saw, with line numbers corrected and unverifiable claims sent back once or dropped.
- 04
Accounts and billing
Better Auth with magic links and optional GitHub sign-in; Paddle for subscriptions with per-country prices; an admin dashboard for users, credits, usage, sales and reviews.
- 05
Website and hosting
A Next.js marketing site with pricing, a model guide and SEO guides. Server, website and downloads run with Docker Compose behind Caddy on a VPS.
- 06
Tests
Vitest for logic and Playwright for the app's flows, running against canned AI answers so tests cost nothing and never touch the network.
Process
A tool built the way it tells you to work
Forge came out of my own workflow with Claude Code: plan first, give the agent rules, and audit before big changes. It turns that habit into a product.
- 1
Plan, then build
Every feature started as a written plan reviewed before any code.
- 2
Feature branches
Cost estimates, SEO guides and support tooling were each built in their own git worktree.
- 3
Screenshots from tests
Product screenshots come from a Playwright script on demo projects, with no network and no AI cost.
- 4
Eat your own cooking
Forge audits its own repo; its findings became the next plans.