Code Analysis Agent

AI-powered code analysis chatbot with two architecture options.

Quick Start

pnpm install
vercel env pull
pnpm dev

Two Architectures

This demo shows two approaches to building AI agents with Vercel infrastructure:

1. Bash-Tool Mode

Path: /api/bash-tool/chat

Simple stateless API pattern using the bash-tool library.

Request → Create/Reuse Sandbox → Run Agent → Stream Response

Pros:

• Simple, easy to understand
• Fast iteration during development
• No workflow complexity
• Works well for single-turn or simple multi-turn

Cons:

• No automatic retries or durability
• Less resilient to failures

2. Workflow Mode (Production-ready)

Path: /api/chat

Durable workflow pattern using Vercel Workflow DevKit.

Request → Start Workflow → Create Sandbox (checkpointed) → Run Agent → Wait for Hook → Resume

Pros:

• Durable: survives crashes, auto-resumes
• Checkpointed: sandbox created exactly once
• Production-ready patterns
• True multi-turn with hook-based resumption

Cons:

• More complex to debug
• Hook system adds latency
• Overkill for simple use cases

Architecture Comparison

Feature	Bash-Tool Mode	Workflow Mode
Complexity	Low	High
Durability	None	Full
Multi-turn	Basic	Native (hooks)
Sandbox persistence	Client-side	Checkpointed
Best for	Demos, prototypes	Production

Tech Stack

• Next.js 16
• Vercel Sandbox - Isolated code execution
• Vercel Workflow - Durable functions (workflow mode)
• AI SDK - Agent orchestration
• bash-tool - Simplified tool creation

Project Structure

├── app/
│   ├── api/
│   │   ├── bash-tool/chat/   # Bash-tool mode endpoint
│   │   ├── chat/             # Workflow mode endpoint
│   │   └── [id]/stream/      # Workflow hook resume
│   └── components/           # React UI
├── lib/
│   ├── bash-tool-agent.ts    # Bash-tool sandbox management
│   ├── config.ts             # Shared config
│   └── system-prompt.ts      # AI prompts
└── workflows/
    ├── custom-bash/          # Workflow definition
    ├── steps/                # Durable tool steps
    ├── tools/                # Tool factory & sandbox
    └── hooks/                # Multi-turn hooks

Usage

1. Select mode (toggle in header)
2. Ask to analyze a repo: "Analyze https://github.com/user/repo"
3. Follow up with questions about the code
4. Sandbox persists

Example Queries

Once a repo is cloned, try these queries to explore the agent's capabilities (comparisions or diffs, understanding complex code, security and quality, code analysis, metrics):

• "What's the high-level architecture?"
• "How is the project organized? What are the entry points?"
• "Map out the data flow from API request to database and back."
• "Count lines of code by file type (ts, tsx, css, etc)."
• "What's the ratio of test code to application code?"
• "Which functions have the most parameters? Flag potential refactoring candidates."
• "How does authentication work? Trace the flow from login to session creation."
• "Explain the error handling strategy - how do errors propagate?"
• "Find all TODO, FIXME, and HACK comments - what technical debt exists?"
• "What middleware runs on each request? In what order?"
• "List all environment variables used across the project."

Links

• Vercel Sandbox Docs
• Vercel Workflow Docs
• AI SDK Docs