Design, build, review, deploy, monitor, and maintain software with intelligent AI agents.
Next.js 15 · React 19 · Fastify · Prisma · PostgreSQL · BullMQ · WebSocket · TypeScript
Forge AI puts the whole software delivery lifecycle behind one application, backed by a catalog of nine specialised AI agents that a human stays in control of.
Most AI coding tools operate on a single file or a single prompt. Forge AI operates on a project — it holds your repository context, tasks, integrations, and the history of every agent run, so an agent asked to review a pull request already knows what the service does.
| Stage | Agents |
|---|---|
| Design | Software Architect, Sprint Planner |
| Build | Documentation Generator |
| Review | Code Review, Pull Request Review, Bug Finder |
| Operate | Deployment Assistant, Log Analysis, Incident Investigation |
Forge AI needs a PostgreSQL database. That is the entire list.
Redis, Kafka, Docker, pgvector, and an AI provider key are all optional. Without them the platform selects in-process implementations — an embedded queue, an in-memory event bus, application-layer vector search, and a simulated model provider — so the whole product works end to end on a laptop. Adding the real services later changes configuration, not code.
GET /ready reports which implementation each subsystem chose:
{
"ready": true,
"checks": { "db": "ok" },
"drivers": {
"queue": "memory", "bus": "memory", "kv": "memory",
"vectorSearch": "fallback", "events": "disabled", "ai": "simulated"
}
}The trade-off is explicit rather than hidden: the in-process drivers are correct for a single node. Running a second node needs Redis, because a browser may be connected to a different node than the worker producing its tokens.
git clone https://github.com/mansiverma897993/Forge-AI.git
cd Forge-AI
pnpm installNo PostgreSQL installed? Start the bundled one — real Postgres binaries from npm, stored in .data/postgres, no installer and no admin rights:
pnpm db:localThen, in another terminal:
pnpm setup # writes .env with generated secrets, applies the schema, seeds demo data
pnpm dev # web on :3000, api on :4000| Web app | http://localhost:3000 |
| Documentation | http://localhost:3000/docs |
| API | http://localhost:4000 |
| API reference (Swagger) | http://localhost:4000/docs |
Sign in with the seeded account — founder@forge.ai / Password123!
Agents run against a simulated provider until you add ANTHROPIC_API_KEY, OPENAI_API_KEY, or GEMINI_API_KEY to .env.
Full documentation lives at /docs in the running app, and as markdown in docs/content/.
| Introduction | What it is and how a run works |
| Quickstart | Running locally in five minutes |
| Configuration | Every variable and what unset means |
| Architecture | Driver seams and the request path |
| Agents and runs | Lifecycle, streaming, cancellation |
| Knowledge and RAG | Retrieval, with and without pgvector |
| Agent catalog | All nine agents and their inputs |
| REST API | Every endpoint |
| Deployment | Vercel + Render, step by step |
| FAQ | Common problems and their real causes |
apps/web Next.js 15 frontend — app, landing page, documentation site
apps/api Fastify backend — REST, WebSocket, SSE, workers
packages/core Shared zod contracts, domain types, crypto, agent catalog
packages/db Prisma schema, migrations, seed
packages/config Shared tsconfig presets
docs/content Documentation, rendered at /docs
infra/ Docker, Kubernetes, Nginx, Prometheus/Grafana/Loki
packages/core is the contract boundary — the API and the web app import their types from it, so a change to a request shape is a type error on both sides rather than a runtime surprise.
- Nine agents, each a typed definition driving the UI form, API validation, and orchestrator from one schema
- Queued execution with bounded concurrency, retry with exponential backoff, and cooperative cancellation
- Realtime streamed output over WebSocket or SSE, plus a live activity feed and notifications
- RAG over per-project knowledge, indexed with pgvector or ranked in the application layer
- Multi-tenant organisations with owner/admin/member/viewer roles enforced in the service layer
- Integrations for GitHub, Slack, Docker, and Kubernetes with signature-verified, de-duplicated webhooks
- Governance — encrypted credentials, append-only audit trail, rotating refresh tokens, rate limiting
- Observability — Prometheus metrics, structured logs, per-subsystem readiness reporting
| Command | Description |
|---|---|
pnpm setup |
Generate .env, apply schema, seed demo data |
pnpm db:local |
Start the bundled PostgreSQL |
pnpm dev |
Run web + api in watch mode |
pnpm build |
Build all packages |
pnpm typecheck |
Type-check the monorepo |
pnpm test |
Run test suites |
pnpm db:push |
Sync the schema without a migration |
pnpm db:seed |
Seed demo data |
pnpm db:studio |
Open Prisma Studio |
pnpm docker:up |
Start the full local stack (Postgres, Redis, Redpanda) |
Both halves fit in free tiers — apps/web on Vercel, apps/api on Render, and any managed Postgres. Redis is only needed once you run more than one node. See the deployment guide.
MIT © 2026 Mansi Verma