hunhee98

hunhee98 / pluck

Public

MCP-native code retrieval for AI agents — 84-88% fewer read tokens, BM25F + semantic search, AST chunks, session dedup

ai-agents bm25 claude-code cli code-intelligence
12
0
94% credibility
Found May 18, 2026 at 33 stars -- GitGems finds repos before they trend. Get early access to the next one.
Sign Up Free
AI Analysis
Rust
AI Summary

Pluck is a local code search engine designed to make AI coding assistants work more efficiently. Instead of having AI agents read entire files repeatedly (wasting context window tokens), pluck organizes code into smart chunks that can be searched by keyword or concept. It integrates directly with AI assistants like Claude Code, Codex, and Cursor through a standard protocol, providing tools for reading code files, searching across a codebase, finding function dependencies, and compressing build logs. The project is written in Rust, published on crates.io, MIT-licensed, and includes comprehensive benchmarks showing 84-88% token savings on code reads and 71% compression on CI logs.

How It Works

1
💡 You hear about a smarter way to code

Your AI coding assistant is wasting time re-reading the same files over and over, burning through your context window.

2
🔧 You install pluck for your project

You run a simple setup command that connects pluck to your AI assistant, making it the default way to read and search code.

3
📚 Your codebase gets organized instantly

Pluck scans your project and builds a lightning-fast index, understanding your code's structure, functions, and relationships.

4
🤖 Your AI assistant discovers your code differently

Instead of dumping entire files, the AI sees smart outlines showing just the important parts - signatures, key functions, and relationships.

5
You search for what you need
💬
Natural language search

You ask 'where is the payment flow handled?' and pluck understands the concept, not just keywords.

⌨️
Exact symbol search

You look for a specific function name and pluck finds it instantly with its exact context.

6
✨ Duplicates disappear from your context

When your AI asks about something it already knows, pluck smartly replaces it with a tiny note instead of repeating the same code.

🎉 Your task completes faster than ever

Your AI assistant finishes the work using far fewer tokens, your build logs are compressed, and everything just works better.

Sign up to see the full architecture

5 more

Sign Up Free

Star Growth

See how this repo grew from 33 to 12 stars Sign Up Free
Repurpose This Repo

Repurpose is a Pro feature

Generate ready-to-use prompts for X threads, LinkedIn posts, blog posts, YouTube scripts, and more -- with full repo context baked in.

Unlock Repurpose
AI-Generated Review

What is pluck?

Pluck is a Rust-based code retrieval engine that replaces cat and grep as the default way AI coding agents read and search code. It exposes symbol-aware code reading through the Model Context Protocol (MCP), using Tree-sitter AST parsing to chunk code intelligently. The core pitch: agents waste massive context window tokens reading the same files repeatedly or scrolling past irrelevant code. Pluck cuts eligible code-read tokens by 84-88% through smart outlines that show signatures and tiny helper bodies, letting agents fetch only the large function bodies they actually need.

Why is it gaining traction?

The token savings hook is real. Session deduplication replaces overlapping code chunks with one-token placeholders, and warm search stays under 0.1ms. The architecture combines BM25 keyword matching with static semantic embeddings (model2vec-style, no transformer inference at runtime) using reciprocal rank fusion. Every tool has a `--raw` fallback that behaves byte-for-byte like cat or grep, so agents lose nothing by defaulting to pluck. The incremental file watcher keeps the index fresh within 171ms of a save.

Who should use this?

AI coding agent power users running Claude Code, Codex, or Cursor who want to reduce token costs and speed up retrieval. Teams benchmarking their agent workflows will find the reproducible benchmark harness valuable. Developers building MCP server integrations for other agents should watch this space, especially as the roadmap targets Aider, OpenHands, Cline, and Continue in v0.9.0.

Verdict

Pluck solves a real problem with a technically sound approach, but at 12 stars and v0.4.0, the credibility score of 0.949999988079071% reflects a project in early growth phase. The gated benchmarks and regression testing infrastructure suggest the team cares about measured claims over marketing. Worth evaluating for agent-heavy workflows, but treat as experimental until v1.0.0 stabilizes the MCP contract.

Sign up to read the full AI review Sign Up Free

Similar repos coming soon.