Devmind

An AI dev assistant that can read your GitHub repos, search the web, and reason about your code — with answers streaming back word by word in real time.

Live demo: https://devmind-mu.vercel.app

What it does

Ask DevMind anything about your code or repos. It decides which tools to use, calls them, and streams the answer back as it generates.

Example prompts:

• Summarise this repo: samnit007/pr-reviewer-agent
• Read the README of samnit007/documind and suggest improvements
• Search the web for LangGraph best practices 2026
• Look at samnit007/devmind and tell me what tests are missing

Screenshots

Streaming response with tool call badges

Multi-tool reasoning

Architecture

React UI (Vite + Tailwind)
      │
      │  POST /api/chat  (SSE stream)
      ▼
FastAPI backend
      │
      ├── Claude API (tool_use + streaming)
      │         │
      │    decides which tools to call
      │         │
      ├── GitHub tool   →  PyGithub → GitHub API
      ├── Web search    →  Tavily API
      └── File reader   →  local filesystem (sandboxed)

MCP Server (stdio transport)
  └── same tools exposed via Model Context Protocol
      → connect via Claude Desktop or MCP Inspector

Key technical details

Claude tool use + streaming together Claude's streaming API and tool use API work differently — streaming delivers text deltas, tool use interrupts the stream for function calls, then resumes. This backend handles both in a single agentic loop: stream text → detect tool call → execute tool → inject result → continue streaming.

MCP server backend/mcp_server.py exposes the same four tools via the official mcp Python SDK over stdio transport. Connect it to Claude Desktop by adding to your claude_desktop_config.json:

{
  "mcpServers": {
    "devmind": {
      "command": "python",
      "args": ["/path/to/devmind/backend/mcp_server.py"]
    }
  }
}

Real-time SSE streaming The backend yields text/event-stream chunks as Claude generates them. The React frontend reads the stream with the Web Streams API and appends each delta to the message in state — no polling, no waiting for the full response.

Tool call badges The frontend shows amber "in progress" badges when a tool is called and green "done" badges when it completes — before the answer text appears. This makes the agent's reasoning visible rather than hiding it behind a spinner.

Stack

Layer	Choice
Frontend	React 19 + TypeScript + Tailwind CSS v4
Backend	Python + FastAPI
AI	Claude Sonnet (tool use + streaming)
Tools	PyGithub, Tavily search, local file reader
MCP	Official mcp Python SDK (stdio transport)
Deploy	Railway (backend) + Vercel (frontend)

Run locally

Backend

cd backend
python3.11 -m venv venv && source venv/bin/activate
pip install -r requirements.txt
cp .env.example .env   # fill in keys
uvicorn app.main:app --reload --port 8000

Frontend

cd frontend
npm install
npm run dev   # http://localhost:5175

MCP server (optional)

cd backend
source venv/bin/activate
python mcp_server.py

Connect via MCP Inspector or Claude Desktop.

Environment variables

Variable	Required	Description
ANTHROPIC_API_KEY	Yes	Anthropic API key
GITHUB_TOKEN	Yes	GitHub PAT (read access)
TAVILY_API_KEY	Yes	Tavily search API key (free tier)
CLAUDE_MODEL	No	Defaults to claude-sonnet-4-6
FILES_ROOT	No	Sandbox root for local file reads (defaults to ~)

What I'd add next

• Conversation memory — persist chat history across sessions (currently stateless per request)
• Code execution tool — run Python snippets in a sandbox and return output
• Streaming to MCP — MCP server currently returns full responses; streaming over SSE transport would improve latency
• Auth — right now anyone with the URL uses your GitHub token and Tavily credits