ntc-code-map

Portable MCP code-map server for token-efficient coding agents.

ntc-code-map builds a lightweight index of your repository's files and symbols, then serves a compact, task-aware code map over the Model Context Protocol (MCP). Coding agents like Codex use it to understand project structure before reading files — saving thousands of tokens per session.

Why?

LLM coding agents waste tokens on broad grep, find, and full-file reads to orient themselves. ntc-code-map compresses this into a single ranked map call:

Quick Start

1. Install

pip install mcp-ntc-road-map-codex-agent

2. Initialize in your project

cd your-project

# Create config
ntc-code-map init

# Build the index
ntc-code-map index

# Check status
ntc-code-map status

3. Use with Codex

# Auto-configure Codex MCP integration
ntc-code-map init-codex

# Auto-inject workflow into AGENTS.md
ntc-code-map init-agents

That's it! Codex will now use ntc-code-map automatically.

MCP Tools

When running as an MCP server, these tools are available:

Recommended agent workflow

1. index_status    → is the index fresh?
2. index_repo      → rebuild if stale
3. repo_map(task)  → get ranked context
4. find_symbols()  → focused follow-up
5. Serena tools    → exact symbol navigation

CLI Reference

ntc-code-map init             # Create .ntc-code-map.toml
ntc-code-map index            # Index source files + symbols
ntc-code-map status           # Show index status
ntc-code-map repo-map "task"  # Generate task-aware map
ntc-code-map find-symbols "query"
ntc-code-map find-files "query"
ntc-code-map module-map src/
ntc-code-map serve            # Start MCP server (stdio)
ntc-code-map init-codex       # Configure Codex MCP
ntc-code-map init-agents      # Inject workflow into AGENTS.md
ntc-code-map doctor           # Check dependencies
ntc-code-map version

Configuration

Create .ntc-code-map.toml in your project root:

[project]
name = "my-project"
root_markers = [".git", "pyproject.toml", "package.json"]

[index]
include_exts = [".py", ".ts", ".js", ".go", ".rs", ".java", ".md"]
ignore_dirs = [".git", "node_modules", "dist", "build", ".venv"]

[ranking]
default_token_budget = 3500
max_files = 80
max_symbols_per_file = 24

How it works

1. Index: Scans source files, extracts symbols using ctags (with regex fallback), stores everything in a local SQLite database (.ntc-code-map/index.db).

2. Score: When repo_map(task) is called, it scores every file against the task description using symbol names, file paths, and content matching.

3. Map: Returns a compact Markdown map with ranked files, their symbols (kind, line, signature), and next-step recommendations — all within the token budget.

Dependencies

• Python ≥ 3.11
• mcp ≥ 1.0.0 (for MCP server)
• Optional: ctags for better symbol extraction (falls back to regex)

Development

git clone https://github.com/ntc-ai/ntc-code-map.git
cd ntc-code-map
pip install -e ".[dev]"
pytest

License

MIT