AgentForge

Bikin AI agent production-ready dalam 60 detik. Input deskripsi → output system prompt + tool definitions + test cases. Export ke Cursor, Hermes Agent, MCP, ChatGPT GPT, OpenAI Functions, Anthropic Tools.

---

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. Powered by Xiaomi MiMo V2.5 (reasoning + structured-output flagship), it converts a 1-3 sentence description into:

• ✅ System prompt — production-ready (400-1500 words), role + behavior + constraints + output format + edge case handling + anti-hallucination clauses
• ✅ Tool definitions — OpenAI/Anthropic/MCP-compatible JSON Schema, only the tools that make sense for the agent
• ✅ Test cases — 4-8 input/expected-behavior pairs covering happy path + edge cases + adversarial inputs
• ✅ Multi-format export — Cursor Rules, Hermes Skill, MCP Server boilerplate (TypeScript), Claude Code prompt, ChatGPT custom GPT, OpenAI Functions JSON, Anthropic Tools JSON

Live demo

👉 agentforge.kleinnn.my.id

Try the presets (Customer Service Bot, Code Reviewer, Trading Sniffer, Hermes Skill Maker, ...) or describe your own agent in plain Indonesian/English.

Stack

• Frontend: Next.js 14 (App Router) + Tailwind CSS + TypeScript
• Reasoning model: Xiaomi MiMo V2.5 Pro (configurable via env)
• API: OpenAI-compatible streaming chat completions
• Deploy: Vercel Edge Runtime (60s max duration)
• Storage: localStorage (no DB, no auth, fully client-side state)

Quickstart (local dev)

Requires Node.js 18+.

git clone https://github.com/agentforge-dev/agentforge.git
cd agentforge
npm install

# Configure model provider (any OpenAI-compatible endpoint works)
cat > .env.local <<EOF
# Xiaomi MiMo V2.5 (recommended)
OPENAI_API_KEY=mimo-XXXXXX
OPENAI_BASE_URL=https://api.xiaomimimo.com/v1
MODEL=xiaomi/mimo-v2.5-pro

# Or OpenAI:
# OPENAI_API_KEY=sk-XXXXXX
# OPENAI_BASE_URL=https://api.openai.com/v1
# MODEL=gpt-4o

# Or Anthropic-compatible (via gateway):
# OPENAI_API_KEY=sk-ant-XXXXXX
# OPENAI_BASE_URL=https://anthropic-gateway.example/v1
# MODEL=claude-sonnet-4.5
EOF

npm run dev
# → http://localhost:3025

Deploy to Vercel

1. Fork this repo
2. Click Deploy to Vercel (or vercel deploy from CLI)
3. Add env vars in Vercel project settings:
   • OPENAI_API_KEY — your MiMo / OpenAI / Anthropic-compat key
   • OPENAI_BASE_URL — endpoint URL (defaults to MiMo)
   • MODEL — model id (defaults to xiaomi/mimo-v2.5-pro)
4. Done. Vercel handles edge runtime + global CDN automatically.

Export formats

Format	File	Use with
Hermes Skill	SKILL.md	Hermes Agent skill registry
Cursor Rules	.cursorrules	Cursor IDE project rules
MCP Server	server.ts	Model Context Protocol server stub
Claude Code	agent.md	Claude Code persona file
ChatGPT GPT	gpt.txt	Custom GPT instructions field
OpenAI Functions	request.json	tools parameter for /v1/chat/completions
Anthropic Tools	request.json	tools parameter for Messages API

Example output

Input: "Code review agent untuk PR Python/TypeScript. Cek security (SQL injection, XSS), performance, code style. Output structured: severity + line + suggestion."

Output (Hermes Skill format, abridged):

---
name: code-reviewer
description: Reviewer PR Python/TypeScript with security/perf/style checks, structured output
---

# CodeReviewer

## System prompt

You are a senior security-aware code reviewer. Analyze diffs in Python or TypeScript.
For each issue, output a structured finding:

  - severity: critical | high | medium | low | info
  - file: path
  - line: line number
  - category: security | performance | style | correctness
  - issue: <description>
  - suggestion: <concrete fix>
  - confidence: 0.0-1.0

Security focus: SQL injection (parameterized queries), XSS (escaping/sanitization),
SSRF, path traversal, hardcoded secrets, unsafe deserialization, ...

## Tools

- `read_file(path: string)` — fetch file content for context
- `run_static_analysis(path: string, lang: 'py'|'ts')` — invoke ruff/eslint
- `git_blame(path: string, line: number)` — see who introduced the line

## Test cases

1. Input: `<diff with SQL string concat>` → flag as critical SQL injection
2. Input: `<clean diff>` → empty findings array
3. Input: `<TS file with eval>` → flag as high security
...

AI tools used to build AgentForge

This project itself was built using AI agent tooling — the same workflow AgentForge optimizes:

• Hermes Agent — workflow orchestration (skills, terminal, file ops)
• Claude Code (Sonnet 4.7) — TypeScript / React authoring
• 9router — local OpenAI-compatible LLM proxy for dev
• MiMo V2.5 — runtime spec generation (will become primary post-grant)
• Cursor — fine-grained code edits
• Vercel — preview deployments
• Tailwind CSS + shadcn/ui patterns — UI

Roadmap

• Multi-format export
• Streaming generation UX
• Live demo on Vercel
• Spec versioning + diff viewer (git-style)
• Auto-evaluator: run testCases against spec on N models, score
• Multi-agent compositions (orchestrator + workers as a graph)
• 1-click deploy: spec → Cloudflare Worker / Vercel function
• Skill marketplace: share + remix specs (community CC0/MIT)

License

MIT — see LICENSE

---

Built with ❤️ by Klein · powered by Xiaomi MiMo V2.5