CrewForge

One local workspace for all your text AI models.

CrewForge is a modern localhost web app for connecting text/code AI providers, discovering models, running server-side health checks, separating free/paid/broken models, selecting a usable crew, and running a planned multi-agent workflow with saved run history. It provides a premium, responsive interface and robust local-first secure storage.

Features

• Multi-Provider Support: Seamlessly connect to G4F, OpenRouter, Gemini, Groq, OpenAI, xAI, Ollama, and Anthropic.
• Two Chat Modes:
  • Direct Chat: Converse with a single AI model.
  • Forge Crew: Automate complex tasks using a multi-agent workflow (Orchestrator → Planner → Workers → Reviewer → Synthesizer).
• Secure by Default: API keys are encrypted and stored locally. They are never sent to the frontend or included in logs.
• Run History: Easily review past workflows, agent outputs, and final results.
• Premium UI: Designed with modern web aesthetics, including glassmorphism, fluid animations, and robust light/dark mode support.

Prerequisites

• Node.js 22.5+ (the server uses the built-in node:sqlite module).
• pnpm 11+ (npm install -g pnpm).

Run Locally

pnpm install
pnpm dev

Open:

http://127.0.0.1:8787

The local server serves both the API and the built React app from the same localhost origin.

Project Structure

crewforge/
├── apps/
│   ├── server/   # Node + Express + TypeScript API, SQLite storage, provider adapters, crew orchestration
│   └── web/      # React + Vite + TypeScript frontend
├── package.json  # pnpm workspace root with dev/build/typecheck scripts
└── pnpm-workspace.yaml

What Is Included

• React + Vite + TypeScript frontend.
• Node + Express + TypeScript backend.
• SQLite local storage using Node's built-in SQLite module.
• Setup, Models, Chat, Runs, and Settings pages.
• Chat page with two usage modes:
  • Chat Mode: send a message directly to one selected model.
  • Forge Crew Mode: manually assign selected models to Orchestrator, Planner, Workers, Reviewer, and Synthesizer.
• Provider adapters for G4F public endpoints, OpenRouter, Gemini, Groq, OpenAI, xAI, Ollama, and Anthropic.
• Local API key storage modes, including encrypted local storage and session-only storage.
• Secret redaction for logs, SSE events, and run storage.
• Model discovery, text/code filtering, health-check queue, health cache, and provider cooldowns.
• Free/no-key/free-tier, paid/key-required, and broken/passive model sections.
• Ready-only model selection.
• SSE live run events.
• Single-model direct chat.
• Manual Forge Crew configuration with 1 to 5 Workers and optional auto-fill.
• Orchestrator, planner, worker, reviewer, and synthesizer flow.
• Active Forge Crew runs survive page navigation; the Chat page reconnects to saved live progress when reopened.
• Auto-fill prefers the selected model with the widest context window for Synthesizer.
• Theme setting supports System, Light, and Dark modes.
• Fallback across selected ready free models.
• Run history with agent outputs, reviewer notes, final answer, and errors.

Chat Modes

Chat Mode uses only one model selected on the Models page:

User -> Selected Model -> Response

Forge Crew Mode uses only models selected on the Models page, then streams progress from:

Orchestrator -> Planner -> Workers -> Reviewer -> Synthesizer

Useful endpoints:

POST /api/chat/single
POST /api/runs
GET /api/runs/:runId/events

Provider Notes

G4F Public Endpoints is experimental. The MVP only uses public no-key style endpoints and does not use cookies, HAR files, browser scraping, private sessions, or g4f.space/v1 as a no-key endpoint.

OpenRouter, Gemini, Groq, OpenAI, xAI, and Anthropic require API keys for practical use. Paid/key-required models are locked by default and are not selected automatically.

Ollama is local-only and is discovered from:

http://127.0.0.1:11434

Security

CrewForge never returns saved API keys to the frontend. Keys are never included in health checks, SSE events, or run history. Health checks use only:

Reply with exactly: OK

Current Limitations

• Paid model use is visible but remains locked for some automated selection paths; automatic paid fallback is experimental.
• Public no-key endpoints can be slow, unavailable, or rate limited depending on external factors.
• OS keychain mode is represented in settings but securely falls back to session-only behavior to avoid cross-platform native dependency friction.

License

Released under the MIT License.