zamalali/DeepGit

What it solves

DeepGit is an agentic research engine designed to find the most relevant GitHub repositories based on a user's full intent rather than simple keyword searches. It solves the problem of "hidden gems" being buried by star counts or lossy keyword-based search results, ensuring that the best-fitting project—regardless of size—is surfaced.

How it works

DeepGit operates as an adaptive agent that escalates effort based on uncertainty. It follows a multi-step process:

1. Intent Analysis: An LLM converts the user's request into a plan containing search angles, hard constraints (e.g., language), and anti-patterns.
2. Broad Gathering: It performs batched GitHub GraphQL searches across keywords and topics.
3. Semantic Recall: It uses a local vector index (LanceDB + fastembed) to pull in near-neighbors from previously seen repositories, allowing recall to improve over time.
4. Prefiltering: It narrows the shortlist using free signals like activity, license, and health.
5. Ranking: An LLM ranks the shortlist based on READMEs and signals.
6. Escalation: If the top picks are contested, the agent reads the actual source code, tests, and manifests to make a final, reflected decision.

Who it’s for

It is intended for developers and researchers who need to find high-quality, open-source software libraries or tools that meet specific technical constraints (e.g., a specific language or hardware limit) without relying solely on GitHub's native search.

Highlights

• Intent-First Search: Preserves full request details as requirements and preferences instead of reducing them to keywords.
• Adaptive Pipeline: Minimizes LLM costs by using a confidence gate to only read source code when necessary.
• Compounding Semantic Recall: Local vector indexing ensures the system gets smarter with every search.
• MCP-Native: Ships as a Model Context Protocol server, allowing AI assistants like Claude Desktop or Cursor to call DeepGit directly.
• Evidence-Based Ranking: Ranks by genuine fit and technical evidence rather than popularity (stars).