jahala/tilth

Smart(er) code reading for humans and AI agents. Reduces cost per correct answer by ~40% on average. Install: cargo install tilth -or- npx tilth

What it solves

It reduces the high cost and inefficiency of AI agents reading code. Instead of making multiple tool calls (like globbing, reading entire files, and grepping) to find a specific function or understand a call chain, tilth provides structural awareness in a single call, significantly lowering token usage and increasing accuracy.

How it works

Tilth combines ripgrep for fast searching, tree-sitter for AST parsing across 16 languages, and memory-mapped file reads for speed. It analyzes code structurally rather than just as text, allowing it to:

  • Outline large files: Automatically provides a structural summary of large files instead of the full content to save tokens.
  • Find definitions: Prioritizes where symbols are defined over where they are simply mentioned.
  • Trace call chains: Includes a "callee footer" in search results so agents can follow function calls without new searches.
  • Structural Diffing: Detects changes at the function level rather than the line level.
  • MCP Integration: Acts as a Model Context Protocol (MCP) server, allowing AI agents (like Claude Code or Cursor) to use these structural tools directly.

Who it’s for

  • AI agents and LLM-based coding assistants that need to navigate large codebases efficiently.
  • Developers who want a fast, structural CLI tool for exploring code definitions, dependencies, and call sites.

Highlights

  • Cost Reduction: Benchmarks show a 38-44% reduction in cost per correct answer across various models.
  • Structural Search: Uses tree-sitter to find actual definitions and call sites, not just string matches.
  • Smart Outlining: Switches between full content and structural outlines based on token count (~6000 tokens).
  • Broad Language Support: Supports 16 languages including Rust, TypeScript, Python, Go, and Java.
  • Hash-Anchored Editing: An optional edit mode that uses hashes to ensure edits are applied to the correct version of the code.

Related

  • Project
    tirth8205/code-review-graph

    A tool that builds a structural graph of your codebase to provide AI assistants with precise context, significantly reducing token usage and improving code review efficiency.

  • Project
    defendend/Claude-ast-index-search

    A structural, AST-aware code navigation CLI that builds a local index of symbols and references to help humans and AI agents navigate large codebases efficiently.

  • Project
    Muvon/octocode

    Octocode is an open‑source Rust tool that builds a live, tree‑sitter‑derived knowledge graph of a codebase and exposes it via an MCP server. AI assistants can perform semantic search, LSP‑level navigation, and graph‑based queries (imports, calls, etc.) across 16 languages. It supports optional vector embeddings for hybrid retrieval, offers a local‑first privacy model, and includes a reproducible benchmark showing strong code‑search performance.

  • Project
    jgravelle/jcodemunch-mcp

    A token-efficient MCP server that uses tree-sitter AST parsing to allow AI agents to retrieve precise source code symbols rather than reading entire files.

  • Project
    cocoindex-io/cocoindex-code

    An AST-based semantic code search tool that enables AI coding agents and developers to find code by meaning and structure, reducing token usage by up to 70%.