Codedict – a no-AI code index that cut my Claude Code tokens by 63%

Developer Marqueos released Codedict, a local code-indexing tool engineered to eliminate repetitive context ingestion in coding assistants without using neural embeddings or machine learning models. According to github.com's report, the utility indexes projects into a local SQLite database using tree-sitter grammars and feeds exact symbol graphs into Claude Code on turn one. Across benchmark tests on a multi-language codebase of roughly 250 files, Codedict reduced total token consumption by 63 percent and turn counts by 75 percent.

This report covers Codedict's architecture, benchmark data, multi-language capabilities, and operational boundaries for software engineers running command-line agent sessions. It is written for developers whose token budgets and usage limits get depleted by iterative file-searching loops. Readers will find verified performance numbers, configuration instructions for Claude Code and Model Context Protocol clients, syntactic parsing rules, and an analysis of how deterministic code graphing compares against traditional retrieval-augmented generation pipelines.

What shipped in Codedict – a no-AI code index that cut my

The core utility is a static analyzer that constructs a deterministic structural map of a codebase using tree-sitter parsers instead of neural embeddings. During an initial scan, Codedict identifies every function, class, interface, HTTP route, and SQL table, recording qualified identifiers, file locations, line numbers, docstrings, decorators, and call hierarchies directly into an SQLite database at dot-codedict slash index dot db. The indexer discovers files using git ls-files, respects gitignore configurations, and automatically excludes minified files, generated artifacts, environmental variable files, and cryptographic certificates.

The system produces two distinct outputs from a single local indexing pass. For artificial intelligence models, Codedict populates compact outlines and prepared context answers in dot-codedict slash ai, accessible through command-line subcommands and a local Model Context Protocol server. For human developers, the tool compiles a comprehensive project manual into dot-codedict slash docs slash README dot md, featuring Mermaid architecture diagrams that trace relationships across modules, database tables, and API routes. The generated dot-codedict folder includes its own gitignore file so local metadata never clutters source control repositories.

What improved in Codedict – a no-AI code index that cut my

Benchmarking across twelve test sessions on a multi-language project with approximately 250 files revealed significant operational gains over standard agent interactions. Conventional Claude Code workflows frequently re-read between 35,000 and 50,000 tokens during each search turn, causing sequential grep calls to consume substantial token allotments. By serving complete symbol locations and call chains on turn one, Codedict collapsed total conversational turns from 138 down to 34, marking a 75 percent drop. This turn reduction lowered total token consumption from 4,543,176 tokens to 1,666,033 tokens, achieving a measured 63 percent decrease.

Financial expenses decreased from 5.76 dollars to 4.27 dollars, yielding a 26 percent net savings across the benchmarked tasks. The dollar reduction was smaller than the token reduction because Claude Code context cache writes represent the costliest billing tier and remain required for active sessions. Accuracy remained uncompromised, maintaining an identical 12 out of 12 success score with automated checks verifying that responses identified the expected functions and source files. Zero-token lookups on open-source repositories answered three of four queries on psf slash requests and three of three on spf13 slash cobra in one search.

What you gain from Codedict – a no-AI code index that cut my

Developers gain predictable context retrieval that avoids the common pitfalls of vector-based search pipelines. Traditional retrieval-augmented generation relies on embeddings to match fuzzy text chunks, often fragmenting syntax and omitting critical call sites located across different project directories. In contrast, Codedict parses exact abstract syntax trees to catalog qualified signatures, decorators, and full call hierarchies leading to entry points. When asking where logic resides or which services trigger a method, the tool delivers exact line numbers, call sites, and source definitions immediately, eliminating the multi-turn exploratory grep cycles that exhaust agent tokens.

The indexer provides broad native support for Python, JavaScript, TypeScript, TSX, C-sharp, Java, and Go through tree-sitter grammars, with regular expressions extracting SQL tables and Razor endpoints. The search engine handles linguistic inflections in both English and Portuguese, mapping queries like parsed to ParseFlags or exclusão to excluir. File tracking relies on SHA-256 hashes, updating only changed files upon saving without triggering token-consuming AI recalculations. Developers can also expand language coverage to over 160 grammars provided by tree-sitter-language-pack by placing custom configuration files into the project dot-codedict slash languages directory.

How to get Codedict – a no-AI code index that cut my

The software requires Python 3.11 or higher and installs into user directory tilde slash dot-codedict without root privileges or altering system paths. Installation scripts configure integration hooks inside tilde slash dot-claude slash settings dot json, back up the prior configuration to settings dot json dot codedict-bak, and append operational instructions to tilde slash dot-claude slash CLAUDE dot md. Manual setup is available through pip, while updating involves repeating the initial installation command. To uninstall, developers invoke the setup module with the remove flag and delete the dot-codedict folder.

curl -fsSL https://raw.githubusercontent.com/Marqueos/codedict/main/install.sh | sh

irm https://raw.githubusercontent.com/Marqueos/codedict/main/install.ps1 | iex

pip install git+https://github.com/Marqueos/codedict
python -m codedict setup

npm install -g @anthropic-ai/claude-code
claude --model fable

/model fable

codedict --root <project> mcp

~/.codedict/venv/bin/python -m codedict setup --remove

Beyond native Claude Code integration, Codedict interfaces with Cursor, Codex, Gemini CLI, OpenCode, and Windsurf through its built-in Model Context Protocol server. Appending the output of codedict agent-snippet into AGENTS dot md connects these coding tools directly to repository symbol tables. Developers working with local models through Ollama or LiteLLM can invoke standalone search commands specifying token limits to insert structured context directly into prompts. The CLI also provides direct inspection utilities such as codedict search, codedict callers, codedict routes, and persistent function annotations.

What to watch after Codedict – a no-AI code index that cut my

Engineers must evaluate the operational boundaries of deterministic indexing and account for fixed caching fees. Every Claude Code session carries a structural base cost of roughly 37,000 cached tokens per conversational turn. Because writing context to cache is the most expensive token tier, reducing turn counts yields a 26 percent cost savings rather than tracking the 63 percent token reduction directly. Furthermore, non-code questions or vague conversational prompts like asking for quality assessments lack identifiable symbols, causing the hook to remain silent so the model operates normally.

Call resolution functions without a compiler, classifying discovered call hierarchies into high, medium, or low confidence based on lexical proximity, variable types, inheritance hierarchies, and partial class definitions. Dynamic invocation, runtime reflection, and complex metaprogramming can introduce structural gaps that static parsers cannot detect without full compilation. Teams adopting Codedict should monitor indexing accuracy on proprietary frameworks, maintain custom ignore patterns inside dot-codedictignore, and verify that incremental SHA-256 change tracking coordinates effectively with active local development branches before standardizing the hooks across organizational repositories.