Skip to main content
← All Projects

Forge

Write a spec, run forge build. AI agents generate the codebase.

Python Multi-Agent Google ADK A2A Protocol CLI
Forge preview

The Problem

Going from idea to working code is slow. You either write boilerplate for hours, or you copy-paste ChatGPT output and spend just as long stitching it together. Single-agent LLM tools hit context limits, hallucinate API contracts, and can’t self-correct.

What Forge Does

Forge is an open-source CLI that turns a markdown spec into a complete, working codebase. Not a template generator — a team of specialized AI agents that reason about architecture, write coordinated code, and review each other’s work.

$ forge new my-app -t web-app
$ nano .forge/spec.md
$ forge build --adk

Agent Pipeline

┌──────────┐     ┌──────────────────┐     ┌───────────────────┐
│  Planner │────▶│  Project Manager │────▶│  Parallel Scheduler│
└──────────┘     └──────────────────┘     └─────────┬─────────┘

                          ┌─────────────────────────┼────────────────────┐
                          │              │          │          │         │
                          ▼              ▼          ▼          ▼         ▼
                     ┌─────────┐  ┌──────────┐ ┌──────┐ ┌────────┐ ┌────────┐
                     │ Backend │  │ Frontend │ │  CI  │ │ Deploy │ │Security│
                     └────┬────┘  └────┬─────┘ └──┬───┘ └───┬────┘ └───┬────┘
                          │            │          │          │          │
                          └────────────┴──────────┴──────────┴──────────┘


                                           ┌──────────┐
                                           │ Reviewer  │
                                           │ routes    │
                                           │ fixes back│
                                           └──────────┘

Eight agents, each specialized. The critical innovation: structured API contracts. The backend agent outputs machine-readable endpoint definitions that the frontend consumes directly. Endpoints always match.

Core Architecture

ComponentPurpose
ArtifactBusThread-safe shared store for all generated code and decisions
ContractRegistryStructured API/model/event contracts extracted from backend output
Parallel SchedulerDependency graph computation, concurrent task execution
Agentic FirewallValidates every file write against path allowlists and content patterns
A2A ProtocolGoogle’s Agent-to-Agent spec for inter-agent communication

Two Build Modes

Classic — dependency-aware parallel scheduler. Independent tasks (backend, CI, deploy) run concurrently. Works with any LLM provider including local Ollama.

ADK — Google’s Agent Development Kit with 7 specialized agents and A2A protocol. Better for complex, multi-service projects.

What Gets Generated

  • Full API routes with error handling and correct HTTP status codes
  • Database models (ActiveRecord, Django ORM, raw SQL)
  • Auth systems (session cookies or JWT)
  • React/Svelte/Next.js components matching backend contracts
  • CI/CD pipelines (GitHub Actions, Dockerfile, docker-compose)
  • Deployment configs (Railway, Render, Vercel, Fly.io)
  • Security audit with vulnerability patches

Supports 14 stacks: Rails, Django, FastAPI, Go, Phoenix, HTMX, React, SvelteKit, Next.js, Express, Flask, Node.js, Python, Elixir. Published as forge-ai on PyPI.