GitHits beta 0.9

Product Introduction

1. Definition: GitHits (beta 0.9) is a specialized context layer and retrieval service for AI coding agents. It functions as a version-aware index and navigation tool that provides AI models (like Claude Code, Cursor, Copilot) with programmatic access to real implementation examples, dependency source code, documentation, and package metadata from the open-source ecosystem. It is delivered as a Model Context Protocol (MCP) server.
2. Core Value Proposition: GitHits exists to eliminate the "missing context" problem that causes AI coding agents to fail. It directly addresses agent hallucinations, costly retry loops, and token waste by grounding AI suggestions in verified, real-world code from the dependencies an application actually uses, ensuring agents can navigate the full software stack beyond the local repository.

Main Features

1. Version-Aware Code Indexing & Navigation: GitHits builds an immutable, commit-pinned index of open-source repositories and packages. The MCP tools (e.g., search, code_read, code_grep) allow agents to perform symbol search, grep code, list files, and read exact line ranges across packages and their dependency graphs. The index is stable for a pinned version (e.g., npm package 0.2.5), ensuring reproducible and accurate context retrieval.
2. Implementation Example Retrieval: Using the get_example tool, GitHits analyzes code, issues, discussions, pull requests, and repository context to distill and return a relevant, actionable code implementation example. This moves beyond raw search results to provide synthesized guidance on how similar problems were solved in open source, directly usable by the agent.
3. Comprehensive Package & Dependency Inspection: Agents can query package metadata, dependency relationships, versions, vulnerabilities (CVEs), changelogs, and release notes using tools like pkg_info and pkg_vulns. It supports a wide range of registries including npm, PyPI, crates.io, Maven Central, NuGet, and others. This enables agents to understand dependency internals, version-specific behavior, and security posture.

Problems Solved

1. Pain Point: Agent Hallucination and Retry Loops. When AI agents encounter undocumented dependency behavior or API edge cases, they lack the necessary context, leading them to generate plausible but incorrect code. This results in guesswork, repeated failures, wasted output tokens, and "derailed agents" that never converge on a working solution.
2. Target Audience: AI-powered coding tool users (developers using Claude Code, Cursor, GitHub Copilot, Codex), and engineers working in ecosystems where AI is less reliable due to thinner documentation (e.g., Go, Rust, C++, Python AI frameworks). It is essential for developers building complex systems with deep, often undocumented, open-source dependencies.
3. Use Cases:
   • Researching and correctly implementing an undocumented API or SDK feature (e.g., finding Azure's TranscribeDefinition).
   • Fixing integration bugs where the root cause lies in a dependency's internal behavior (e.g., TanStack Query's persisted mutation defaults).
   • Writing correct code for low-documentation libraries (e.g., CUDA kernels, DuckDB C++ APIs).
   • Conducting dependency risk assessment by inspecting package vulnerabilities and license compliance during development.

Unique Advantages

1. Differentiation: Unlike generic web search or code search engines, GitHits is purpose-built as a discovery engine for AI agents. It provides structured, version-specific, and navigable context directly via MCP tools, not just links. It focuses on finding implementation patterns from the actual codebase of a dependency, not just its documentation page. It complements local repo access by covering the previously "invisible" 70% of the stack.
2. Key Innovation: The core innovation is its version-aware code graph index that maps files, symbols, imports, exports, and dependency relationships at a specific immutable commit. This allows for precise, reproducible code navigation and analysis that matches the exact version used by the agent's project, preventing context drift and ensuring suggestions are grounded in the correct version of the source code.

Frequently Asked Questions (FAQ)

1. How does GitHits work without accessing my private code? GitHits uses your GitHub authentication only to search public open-source repositories. It does not index, search, or access your private repositories. Its MCP server provides external context (from OSS dependencies) to your local coding agent, which already has access to your private codebase.
2. What AI coding tools are compatible with GitHits? GitHits is designed as an MCP server compatible with any AI coding tool that supports the Model Context Protocol. It is explicitly mentioned as being useful with tools like Claude Code, Cursor, and GitHub Copilot. You can start by running npx githits@latest init, which detects your tool and configures it.
3. Does GitHits train on my code or prompts? No. GitHits explicitly states it does not use customer code, prompts, or outputs to train AI models. Data is processed solely to provide the service and is governed by their Privacy Policy and Data Processing Agreement.
4. How does GitHits handle open-source licenses and avoid risk? GitHits includes strict-mode filtering by default to exclude repositories with copyleft or undeclared licenses. It provides license information via package inspection tools. While it minimizes risk by retrieving structured code instead of browsing arbitrary web pages, developers are still responsible for reviewing any generated code before use.
5. Is the GitHits index static? For a pinned package version or commit, the index is static and reproducible—the underlying source code for that version never changes. However, GitHits may re-index that source to improve its internal code graph and parsing. References to moving targets like main or HEAD will track the latest commit.