Developer Intelligence POC

A local proof-of-concept that builds a living knowledge graph from a Go codebase. Every file gets LLM-generated documentation. Every relationship gets documented. Everything stays current on every merge. Query it from Claude Code via MCP.

Quick Start

cd dev-intel-poc
uv sync                         # install deps
uv run python ingest.py         # clone echo, parse, generate docs (~15-20 min)
claude --mcp-config .mcp.json   # start Claude Code with the knowledge graph

Then ask Claude Code:

• "How does routing work in echo?"
• "What files depend on context.go?"
• "Give me an overview of this project"

Prerequisites

• Python 3.11+
• uv (curl -LsSf https://astral.sh/uv/install.sh | sh)
• Ollama running at 192.168.86.172:11434 with qwen2.5:7b
• Claude Code CLI (claude)
• git

Demo Walkthrough

Act 1: "Here's what your codebase knows about itself"

After python ingest.py completes, start Claude Code:

claude --mcp-config .mcp.json

Ask it:

> What does echo.go do?
> How does echo.go interact with router.go?
> Give me an overview of the whole project
> What files depend on context.go?
> Search for anything related to "middleware"

Every answer comes from LLM-generated documentation stored in the knowledge graph — not from reading raw source code.

Act 2: "Now someone pushes a change"

In another terminal:

python simulate_merge.py echo.go

This:

1. Regenerates echo.go's documentation (reflects the new code)
2. Marks all relationships involving echo.go as STALE
3. Marks the repo summary as STALE

Back in Claude Code:

> What does echo.go do?          # fresh doc — mentions the new tracing feature
> What's the repo overview?       # shows [STALE] — knows it's outdated
> Show me all stale docs          # lists everything that needs refresh

Act 3: "The system heals itself"

python refresh_stale.py

Back in Claude Code:

> What's the repo overview?       # fresh again — rewritten to include new capabilities
> Show me all stale docs          # "Everything is fresh!"

Architecture

ingest.py ──→ repos/target/ (git clone)
    │              │
    │         parser.py (tree-sitter AST)
    │              │
    │         docgen.py (Ollama qwen2.5:7b)
    │              │
    └──────→ devintel.db (SQLite)
                   │
              mcp_server.py ──→ Claude Code

No Docker. No external databases. One SQLite file. One MCP server.

MCP Tools

Tool	What it does
get_file_doc(path)	Read a file's generated documentation
get_relationship(a, b)	How two files interact
get_repo_overview()	Project-level summary
get_dependents(path)	What breaks if you change this file
get_dependencies(path)	What this file depends on
search_docs(query)	Keyword search across all docs
get_stale_docs()	List outdated documentation
get_graph_stats()	File count, relationship count, staleness

Project Structure

dev-intel-poc/
├── requirements.txt       # Python deps
├── .mcp.json              # Claude Code MCP config
├── ingest.py              # Main ingestion pipeline
├── simulate_merge.py      # Simulate a code change
├── refresh_stale.py       # Refresh stale docs
├── db.py                  # SQLite backend
├── parser.py              # tree-sitter Go AST parser
├── docgen.py              # Ollama doc generation
├── mcp_server.py          # MCP server for Claude Code
└── devintel.db            # Generated — the knowledge graph

Configuration

Env Variable	Default	Description
LLM_BACKEND	ollama	ollama or openai (for Kiro gateway, OpenRouter, etc.)
OLLAMA_URL	http://192.168.86.172:11434	Ollama endpoint
OLLAMA_MODEL	qwen2.5:7b	Ollama model
OPENAI_URL	http://192.168.86.11:8000	OpenAI-compatible endpoint (Kiro gateway)
OPENAI_MODEL	claude-haiku-4	Model name for OpenAI-compatible API
OPENAI_API_KEY	not-needed	API key (if required by endpoint)
TARGET_REPO	https://github.com/labstack/echo.git	Repo to ingest
MAX_CONCURRENT	4	Parallel LLM requests
DEVINTEL_DB	./devintel.db	SQLite database path
REPO_DIR	./repos/target	Cloned repo location