⛓ ChainForge

Your AI coding team inside VS Code — chat, auto-edit, and self-healing code, powered by 400+ models via one OpenRouter key.

ChainForge isn't just a chat sidebar. It's a small pipeline of AI agents that talk to each other: a router picks the right model for the job, a fallback chain kicks in the moment one fails, and a Denetmen ↔ Postacı ("Inspector ↔ Postman") loop can scan your codebase, find bugs, and fix them automatically — all without you writing a single line of YAML.

✨ What it actually does

• 💬 Project-aware chat — reads your open file and the most relevant files in your workspace automatically, no need to paste code in. Remembers conversation history per project.
• 🛠 Agent mode (auto-edit) — describe a task in plain language; ChainForge plans it, researches if needed, writes the code, shows you a diff, and applies it on approval. Full operation history kept per project.
• 🔎 Denetmen (Inspector) — scans your workspace for bugs, security issues, and code smells. Incremental: after the first full pass, only re-checks files that changed.
• ✉️ Postacı (Postman) auto-fix loop — reads the Inspector's report and sends each broken file to an AI to fix, then re-applies the fix with your approval. Ask ChainForge to check your code, then let it clean up after itself.
• 🧭 Smart routing & fallback — tasks go to the model best suited for them; if a model fails, hits a rate limit, or the format breaks, the chain automatically moves to the next one. Nothing dead-ends.
• ⏱ Live progress + cancel — see every step as it happens ("trying model X…", "planning…", "writing code…") and cancel a running chat or agent task mid-flight.
• 🌐 400+ Models — Claude, GPT, Gemini, DeepSeek, MiniMax, GLM, Qwen, Grok and more, all through your own OpenRouter key.
• 🆓 Works without any paid key — a free-model fallback chain means you can chat and run basic agent tasks with zero cost, zero setup beyond installing.
• 💰 Cost tracking — every AI call's token usage and estimated cost (live OpenRouter pricing) is logged; see a monthly breakdown by model right in the panel.
• 🌍 7 Languages — EN, TR, DE, FR, ES, JA, ZH — full interface and AI response language.
• 🔒 Privacy-first telemetry — fully opt-in; if enabled, only anonymous feature/error data is collected — never your code, prompts, or API key.
• ✂️ Cost-optimized editing — for existing files, Postacı sends the coding model only the relevant code block instead of the whole file (verified byte-for-byte against the source, with an automatic full-file fallback whenever it can't be verified). Measured 78% lower cost per task on a real 400+ line file.
• 📑 Full-tab view — open ChainForge in a large editor tab instead of the sidebar, when you want more room.
• 🖱 Right-click quick tasks — select code (or just place your cursor in a file) → right-click → "ChainForge: Run Task on Selection" — no need to open the panel first.
• 💡 Quick-fix on errors — a lightbulb on any compiler/linter diagnostic offers "Fix with ChainForge", scoped to that one error.
• 🔄 Self-healing free tier — the free-model list is checked against OpenRouter's live catalog every 24h, so it can't silently go stale.

🚀 Quick Start

1. Install ChainForge from the VS Code Marketplace or Open VSX.
2. Open the ChainForge icon in the Activity Bar.
3. Start chatting immediately with free models — no API key required.
4. Want more power? Get a free key at openrouter.ai/keys and paste it into the Settings tab to unlock paid models and agent tasks.
5. Try ChainForge: Agent Task (Auto Edit Files) from the Command Palette to have it write and apply code directly to your project.
6. Try ChainForge: Inspect Files (Denetmen) to scan your codebase, then let the Postacı auto-fix loop clean up what it finds.

🔗 How It Works

Your task → Router picks best agent → (fails?) → Fallback agent → (still stuck?) → Supervisor

Each agent has a model, a role (coding / math / routing / supervisor / fallback / custom), a retry count, and an optional fallback chain — all configurable through the visual panel, no config file editing required (though .chainforge.json is there if you want it).

The Denetmen ↔ Postacı loop works on top of this: Denetmen finds problems, Postacı asks the coding agent to fix them, and the fix goes through the same diff-review-apply flow as any other agent change.

⚙️ Configuration (optional)

ChainForge stores its agent/task setup in .chainforge.json in your workspace root — edit it directly, or use the Visual Editor in the panel.

{
  "agents": {
    "worker": {
      "name": "AI Worker",
      "model": "minimax/minimax-m3",
      "role": "coding",
      "fallback": "fallback",
      "maxRetries": 3
    },
    "fallback": {
      "name": "Fallback",
      "model": "google/gemini-2.5-flash",
      "role": "fallback",
      "fallback": "supervisor",
      "maxRetries": 2
    },
    "supervisor": {
      "name": "Supervisor",
      "model": "anthropic/claude-sonnet-4.5",
      "role": "supervisor",
      "maxRetries": 1
    }
  },
  "tasks": {
    "coding": { "primary": "worker", "description": "Code writing, editing" },
    "review": { "primary": "supervisor", "description": "Code review" },
    "general": { "primary": "worker", "description": "General questions" }
  }
}

💎 Pro — $9.99 one-time

No subscription, no recurring charge — pay once, keep it forever. ChainForge is built and maintained by a single indie developer, so every Pro purchase goes straight into keeping it alive and improving. If you've bought Pro — thank you, genuinely. It means a lot. 💙

Get Pro — $9.99 one-time →

🔐 Security & Privacy

• Path traversal and prompt-injection–resistant file writes: agent-generated file paths are validated before anything touches disk — a task can only create or edit files inside your workspace.
• Input validation on every user-editable field (agent names, models, prompts) — no script injection, no malformed model IDs.
• Telemetry is opt-in, sanitized, and never includes your code, prompts, file paths, or API key.

🤝 Contributing

Issues and PRs welcome on GitHub.

📄 License

MIT — see LICENSE