Forge

Overview

Forge is an AI agent orchestration platform that goes beyond chatbots. Users build visual DAG workflows (blueprints) combining deterministic code nodes with LLM-powered agent nodes, then execute them with real-time SSE streaming. The platform supports GUI automation, terminal orchestration, multi-model providers, knowledge base RAG, eval frameworks, and cross-platform computer use across macOS, Linux, and Windows.

Core Capabilities

• Visual Blueprint Editor — Drag-and-drop DAG builder with 44 node types across 9 categories. Topological execution engine with concurrent layer resolution and context assembly.
• Computer Use (GUI + Terminal) — 12 GUI automation nodes (screenshot, OCR, click, type, hotkey, scroll, drag, focus, find, wait, clipboard, apps) and 6 terminal nodes (session, run, send, logs, poll, fanout). Cross-platform: Steer CLI on macOS, xdotool on Linux, pyautogui on Windows.
• Agent-on-Agent Orchestration — Spawn and control external coding agents (Claude Code, Codex CLI, Gemini CLI, Aider) as workers in tmux sessions. Full lifecycle: spawn, prompt, monitor, wait, capture, stop.
• Multi-Machine Dispatch — Route nodes to different execution targets with capability-based routing. Target registry with health checks and aggregated capabilities.
• Multi-Model Providers — OpenAI, Anthropic, Google APIs with per-node model selection, health monitoring, and model comparison.
• Knowledge Base + RAG — Document collections with chunked upload, semantic search, and knowledge_retrieval blueprint node.
• Eval Framework — 5 grading methods including screenshot_match and ocr_contains for computer use verification.
• Human-in-the-Loop — Approval gate nodes pause execution for human review with approve/reject workflows.
• Safety & Security — App blocklist, command blocklist, rate limiting (30 actions/min), approval gates, and full audit logging for all computer use actions.
• Observability — Distributed traces for all executions, prompt versioning with diff/rollback, cost analytics with per-model breakdowns.

Scale

• 44 blueprint node types across 9 categories
• 13 pre-built blueprint templates
• 3 LLM providers (OpenAI, Anthropic, Google)
• 3 platform targets (macOS, Linux, Windows)
• 20+ CLI command groups
• 17 database migrations with RLS
• 515 backend tests + 21 frontend tests
• 20+ API route modules

Architecture

Next.js 14 frontend with React Flow blueprint editor, consuming a FastAPI backend. The blueprint engine runs DAGs with topological sorting, concurrent layer execution, and context assembly with token budgets. Computer use nodes dispatch to platform-specific executors (Steer/xdotool/pyautogui for GUI, Drive/tmux/PowerShell for terminals). Multi-machine dispatch routes nodes to registered execution targets. Agent-on-agent orchestration spawns external coding agents in tmux sessions via the Drive layer. Supabase provides auth and PostgreSQL with RLS. Everything streams via SSE.

Version History

• v1.9 — Agent-on-agent orchestration, multi-machine dispatch, screen recording, Linux & Windows computer use, cross-platform unification
• v1.8 — Computer use extension (Steer + Drive), safety/security, remote execution, capability detection
• v1.7 — Workflow marketplace, team features, organization RBAC
• v1.6 — Knowledge base + RAG with semantic search
• v1.5 — Observability traces + prompt versioning
• v1.4 — Eval framework + human-in-the-loop
• v1.3 — MCP integration + event triggers
• v1.2 — Multi-model provider system
• v1.1 — Blueprint system with visual DAG editor
• v1.0 — Production release with demo mode and landing page