Enterprise-grade AI Agent System for Frontend Engineering — Constrained by SDD, Powered by MCP for Controlled Perception and Execution, with built-in Hybrid RAG, Hallucination Prevention, Cross-Session Memory, and Skill Management.
FrontAgent is a domain-specific AI Agent system for frontend engineering, featuring a two-stage architecture (Planner + Executor) with SDD (Specification Driven Development) as a hard constraint layer to ensure generated code conforms to project specifications. The system manages three tool capabilities—file operations, browser interaction (via Playwright), and terminal execution—through the MCP protocol for controlled perception and execution.
On the knowledge side, FrontAgent embeds a remote hybrid RAG engine supporting parallel weighted fusion of BM25 keyword retrieval and semantic retrieval, combined with LLM query rewriting and Cross-Encoder reranking (compatible with Jina/Cohere), with Weaviate as an optional vector store and cache export/import workflows. For code hallucination prevention, it provides multi-layer hallucination detection and automatic import/export path validation. The cross-session memory system covers four stages: pre-loading, runtime recall, post-task persistence, and structured storage, persisting project facts, error solutions, and dependency states.
At the execution level, it features phased self-repair: each phase can automatically analyze errors and generate fix steps, with module reference integrity checks upon phase completion. The Skill Lab module provides a complete local skill iteration workflow (scaffold/eval/benchmark/improve/promote). An optional LangGraph graph execution engine supports checkpoints, and the repository management stage automates the full git commit/push/PR workflow. Built with TypeScript, managed via pnpm monorepo + Turborepo, distributed through npm, currently at v0.1.6 under the MIT license.
Core Architecture#
User Input → Agent Core → Output
│
┌───────────┼───────────┐
▼ ▼ ▼
SDD Layer Planner Executor
(Constraints)(Stage 1) (Stage 2)
│ │ │
└───────────┼───────────┘
▼
MCP Layer
┌──────┬────────┬──────┐
│ File │ Web │Shell │
└──────┴────────┴──────┘
Key CLI Commands#
| Command | Description |
|---|---|
frontagent init | Initialize SDD |
frontagent run "<task>" | Execute a task |
frontagent skill list/scaffold/init-evals/benchmark/improve/promote | Skill Lab workflow |
frontagent rag export/import | RAG cache export/import |
Quick Start#
npm install -g frontagent
export PROVIDER="openai"
export BASE_URL="https://api.openai.com/v1"
export MODEL="gpt-4"
export API_KEY="sk-..."
cd your-project
frontagent init
frontagent run "Create a user login page"
Prerequisites: Node.js environment, external LLM API key (OpenAI or Anthropic). If using Weaviate as vector store, a separate Weaviate instance is required and configured via environment variables.