← Back

About AgentForge

AgentForge is an open-source AI agent spec generator. You describe an agent in plain language, AgentForge generates a production-ready system prompt, tool definitions, and test cases. Export to your favorite agent runtime in one click.

Why?

The hardest part of building agents isn't the framework — it's the spec. A shallow system prompt loses to a well-structured one by 40-60% on real benchmarks. Most builders skip this step because writing rigorous prompts + tool schemas takes hours. AgentForge does it in 60 seconds.

How it works

  1. You describe the agent in 1-3 sentences.
  2. A reasoning model (default: Xiaomi MiMo V2.5) generates a structured spec: name, system prompt, tool defs, test cases.
  3. Spec exports to: Cursor Rules, Hermes Skill, MCP Server boilerplate, Claude Code, ChatGPT GPT, OpenAI Functions, Anthropic Tools.
  4. Copy/download → drop into your agent runtime → ship.

Why MiMo V2.5

MiMo V2.5 is Xiaomi's reasoning + multimodal flagship. Three properties make it ideal for spec generation:

  • Strong structured output — strict JSON mode, low schema drift, clean for downstream parsing.
  • OpenAI-compatible API — drop-in for any OpenAI SDK / fetch.
  • Reasoning-tuned — designs better tool sets than non-reasoning models because it actually thinks about each capability separately.

The app supports any OpenAI-compatible endpoint via env vars (OpenAI, Anthropic, local 9router, etc.) — but defaults to MiMo V2.5 because it ships best results for this workload.

AI tools used to build this

  • Hermes Agent — orchestration (skill system, multi-tool flows)
  • Claude Code (Claude Sonnet 4.7) — TypeScript / React authoring
  • OpenClaw — Solana SPL token minting (sister project)
  • 9router — local OpenAI-compatible LLM proxy for dev
  • MiMo V2.5 — runtime spec generation (the model behind the magic)
  • Next.js 14, Tailwind, Vercel for deploy

Roadmap

  • Spec versioning + diff viewer (track agent prompt evolution like git)
  • Auto-evaluator: run testCases against the generated spec on N models, score
  • Multi-agent compositions (orchestrator + workers, output as a graph)
  • Direct deploy: 1-click to Cloudflare Workers / Vercel function
  • Skill marketplace: share + remix specs (MIT license)

Open source

MIT licensed. github.com/agentforge-dev/agentforge

Built with Xiaomi MiMo V2.5 + Hermes Agent + Claude Code · MIT licensed