cyberchitta/llm-context.py

What it solves

Getting the right amount of code context into an LLM is often a manual, friction-heavy process. Users typically either send too much data, hitting token limits, or too little, causing the LLM to miss critical information and require multiple follow-up requests for missing files.

How it works

The tool allows coding agents to curate their own context through a system of composable rules. An agent uses the lc-init command to install a "skill" that teaches it how to identify necessary files, write selection rules in YAML, and verify those rules using a code graph to find missing symbols. The tool can then pack this context—consisting of full files, structural excerpts (signatures and definitions), and filtered file listings—and deliver it via the clipboard, a pipe to another agent, or an MCP (Model Context Protocol) server.

Who it’s for

Developers using AI coding agents (like Claude Code) or LLM chats who need a precise, automated way to manage the codebase context provided to the model.

Highlights

• Agent-led curation: Agents can author and verify their own context rules without manual user intervention.
• Composable rules: Uses a rule system (categorized by prefixes like prm-, flt-, exc-) to define how files are selected and excerpted.
• Context packing: Combines full file contents with structural excerpts to maximize token efficiency.
• MCP Support: Integrates via Model Context Protocol to allow models to pull missing or changed files dynamically.
• Code graph verification: lc-preview helps agents identify symbols defined in files that were accidentally omitted from the selection.