DevMind — AI-Powered Personal Dev Assistant

Ask questions like you're talking to a senior developer sitting next to you. DevMind reads your actual code, searches all of GitHub, and browses the live web to answer.

Built with OpenAI Agents SDK · MCP Servers · Multi-Agent Handoffs · SQLite Memory

---

What It Does

You type a question. DevMind figures out what you need and routes it to the right specialist agent — which then uses real tools (MCP servers) to fetch actual data before answering.

You:      "review src/devmind/guardrails/input_guards.py"

DevMind:  [handoff → Code Review Agent]
          → reads your file via Filesystem MCP
          → searches GitHub for similar guardrail patterns
          → "Line 58: the empty input check uses < 3 chars which
             might be too aggressive for short commands like 'ls'..."

---

Architecture

Your Input
    │
    ▼
Triage Agent (GPT-4o-mini)         ← 3 input guardrails
    │
    ├── file / code question   →   Code Review Agent (GPT-4o)
    │                               ├── Filesystem MCP  (reads local files)
    │                               └── GitHub MCP      (finds real examples)
    │
    ├── library / repo question →  Repo Search Agent (GPT-4o)
    │                               ├── GitHub MCP      (searches repos)
    │                               └── Filesystem MCP  (saves examples)
    │
    └── error / concept / docs →   Web Research Agent (GPT-4o)
                                    ├── DuckDuckGo tool (live web search)
                                    └── Filesystem MCP  (saves reports)
    │
    ▼
Output guardrails (PII + length)
    │
    ▼
SQLiteSession → saves every turn to data/conversations.db

---

Agents

Agent	Model	Tools	Job
Triage Agent	GPT-4o-mini	—	Routes every request to the right specialist
Code Review Agent	GPT-4o	Filesystem MCP + GitHub MCP	Reads your files, reviews code, finds bugs
Repo Search Agent	GPT-4o	GitHub MCP + Filesystem MCP	Searches GitHub, reads real READMEs and source code
Web Research Agent	GPT-4o	DuckDuckGo + Filesystem MCP	Searches the live web, finds docs and solutions

---

MCP Servers

No tool functions written. Agents auto-discover tools from MCP servers at runtime.

MCP Server	How installed	Tools available to agents
@modelcontextprotocol/server-filesystem	npx -y (auto)	read_file, write_file, list_directory, search_files
@modelcontextprotocol/server-github	npx -y (auto)	search_repositories, search_code, get_file_contents, list_commits, get_issues

Web search uses the duckduckgo-search Python package — no API key, no MCP server needed.

---

Guardrails

Guard	Type	Blocks
empty_input	Input	Blank or near-empty messages
jailbreak	Input	"ignore instructions", "DAN mode", etc.
path_traversal	Input	Attempts to read /etc/passwd, ~/.ssh, .env files
pii	Output	Emails, phone numbers, API keys in responses
length	Output	Responses over 2000 characters

---

Memory

Every conversation turn is saved to data/conversations.db via SQLiteSession.

Turn 1: "review my auth file"      → found 2 issues
Turn 2: "fix the first issue"      → knows which issue you mean
Turn 3: "is that a common bug?"    → knows what "that" refers to
Turn 4: "save a full report"       → has the complete context

---

Try These Questions

Code Review

review the file src/devmind/guardrails/input_guards.py and explain what each guard does

Repo Search

find the best open source Python project that uses multi-agent AI and show me a real code example

Web Research

what is the difference between openai agents sdk and langgraph? which one should a beginner use in 2025?

---

Setup

Requirements

• Python 3.11+
• Node.js 18+ (for npx — MCP servers run as subprocesses)
• OpenAI API key
• GitHub Personal Access Token

Install

git clone https://github.com/YOUR_USERNAME/ai-powered-dev-assistant
cd ai-powered-dev-assistant

python3 -m venv venv
source venv/bin/activate      # Windows: venv\Scripts\activate

pip install -r requirements.txt

Configure

Create a .env file in the project root:

OPENAI_API_KEY=sk-...
GITHUB_PERSONAL_ACCESS_TOKEN=ghp_...
FS_ROOT_PATH=/Users/your-username
SESSION_ID=devmind-user
DB_PATH=data/conversations.db

Get your keys:

• OpenAI: platform.openai.com → API Keys
• GitHub PAT: github.com/settings/tokens — scopes: repo, read:org

Run

python -m src.devmind.main

The first run downloads MCP server packages via npx (takes ~30 seconds). Subsequent runs are instant.

---

Project Structure

ai-powered-dev-assistant/
├── README.md
├── CLAUDE.md
├── requirements.txt
├── .gitignore
├── assets/
│   └── screenshots/             ← place demo screenshots here
├── data/                        ← SQLite session DB (gitignored)
└── src/
    └── devmind/
        ├── main.py              ← CLI entry point
        ├── config.py            ← env vars + constants
        ├── agents/
        │   ├── triage_agent.py
        │   ├── code_review_agent.py
        │   ├── repo_search_agent.py
        │   └── web_research_agent.py
        ├── guardrails/
        │   ├── input_guards.py  ← empty, jailbreak, path_traversal
        │   └── output_guards.py ← pii, length
        ├── mcp/
        │   └── servers.py       ← MCPServerStdio factory functions
        ├── tools/
        │   └── web_search.py    ← DuckDuckGo function tool
        └── session/
            └── memory.py        ← SQLiteSession factory

---

SDK Concepts Demonstrated

Concept	Where
MCPServerStdio	mcp/servers.py — connect agents to MCP servers over stdio
Multi-MCP per agent	code_review_agent.py — Filesystem + GitHub on one agent
Tool auto-discovery	Agents discover MCP tools at runtime, no manual registration
handoff() with input_type	triage_agent.py — structured metadata on every handoff
@input_guardrail	guardrails/input_guards.py — tripwire pattern
@output_guardrail	guardrails/output_guards.py — tripwire pattern
SQLiteSession	session/memory.py — persistent cross-turn memory
@function_tool	tools/web_search.py — custom Python tool alongside MCP

---

Built with OpenAI Agents SDK