GitHits – Context layer for AI coding agents

TL;DR

TL;DR: GitHits is an open-source CLI and MCP server that lets AI coding agents query indexed public open-source code, package docs, and changelogs — giving them source-backed evidence beyond model memory or local repo context.

Source and Accuracy Notes

• Project page: githits.com
• Source repository: github.com/githits-com/githits-cli
• License: Apache-2.0 (verified via GitHub API)
• HN launch thread: news.ycombinator.com/item?id=48551593

What Is GitHits?

GitHits is a version-aware open-source context layer for AI coding agents. It runs a local MCP server that agents can query for factual, source-cited evidence from the broader open-source ecosystem — without relying solely on model memory or your local repository.

The core problem it solves: AI coding agents are great at generating code but often hallucinate or guess when working with unfamiliar packages, rare error patterns, or third-party APIs. GitHits gives them a retrieval layer over indexed public open-source code so they can verify behavior before editing.

Setup Workflow

Step 1: Install via npm

npx githits@latest init
```bash

The `init` command signs you in, detects supported coding tools, and configures the local GitHits MCP server.

### Step 2: Automatic Tool Detection

`init` auto-detects and configures the MCP server for these tools:

- Claude Code, Cursor, Windsurf, VS Code / Copilot, Cline
- Claude Desktop, Codex CLI, Pi, Gemini CLI
- Google Antigravity, OpenCode, Hermes Agent, Zed
- Junie, Qwen Code, Kiro, Kilo Code, Factory Droid, Amazon Q CLI

### Step 3: Use in Your Agent

After setup, agents call GitHits automatically when they need source-backed context. You can also prompt them directly:

```bash
Use GitHits Code Navigation to inspect npm:express. Find how middleware
errors are handled, read the relevant source, and explain the fix before
editing code.
```bash

## Deeper Analysis

### MCP Tools Available

GitHits exposes these MCP tools to your coding agent:

| Tool | What it does |
|---|---|
| `get_example` | Find real-world code examples across repositories |
| `search_language` | Search by programming language |
| `search` | Full-text search across indexed code and docs |
| `code_read` | Read exact source files |
| `docs_list` / `docs_read` | Access package documentation |
| `pkg_info` | Inspect package metadata, health, and stats |
| `pkg_vulns` | Check known vulnerabilities |
| `pkg_deps` | Inspect dependency trees |
| `pkg_changelog` | Read version-by-version changelogs |
| `pkg_upgrade_review` | Compare dependency upgrades with evidence |

### CLI Commands

All MCP tools map to CLI equivalents:

```bash
githits example "HTTP retries with exponential backoff in Python"
githits search "router middleware" --in npm:express
githits pkg info npm:express
githits pkg vulns npm:lodash
githits pkg changelog npm:axios

Practical Evaluation Checklist

• Installs via npx with no build step required
• MCP server starts locally and is accessible to configured tools
• Tool detection covers the major AI coding agents (Claude Code, Cursor, etc.)
• Searches return source-cited results with exact file references
• Changelog and upgrade review provide version-specific evidence
• Works entirely locally — no cloud dependency for core functionality

Security Notes

• The MCP server runs locally; no external API calls are required for core retrieval
• Agents never directly access package registry credentials through GitHits
• Package vulnerability data is fetched from public advisory databases

FAQ

Q: Does GitHits replace model knowledge? A: No. It supplements model knowledge with retrievable, source-cited evidence. Agents still use their trained knowledge but can verify against actual source when needed.

Q: Is there a hosted/cloud version? A: Yes at githits.com, but the CLI and MCP server run fully locally and are the primary open-source offering.

Q: What package registries are indexed? A: The index covers public open-source repositories, npm packages, and their associated documentation and changelogs.

Q: How does it differ from Context7 or similar context engines? A: GitHits focuses specifically on the coding agent use case, with built-in MCP tool definitions that map to both a CLI and an agent-facing search interface.

Conclusion

GitHits solves a real problem in the AI coding agent workflow: agents guessing at third-party APIs instead of verifying against source. By packaging a local MCP server with indexed open-source code, it gives agents a retrieval layer that works entirely offline for the core use case. If your team uses Claude Code, Cursor, or similar tools, it is worth running npx githits@latest init and letting your agents use it automatically.