CodeParse

CodeParse is a codebase indexer that gives agents fast, accurate, and compact context on your Python codebases while reducing token usage. It parses your codebase and stores the relationships of your code in a local SQLite database.

See the Benchmark Evaluation section for more details.

Install MCP Server and Skill

1. Put the codeparse binary on your PATH (release asset or make compile → dist/codeparse).
2. codeparse install

Tools

Tool	Description
get_project_overview	Return the largest directories and the project entry points.
get_directory_tree	Return the directory tree.
get_file_symbols	Return the symbol tree of one file.
get_file_imports	Return the import statements of one file.
get_symbol	Return one symbol's definition and its reference counts.
get_symbol_at_lines	Return the symbol that covers a line range in a file.
get_symbol_references	Return the call, access, and type-annotation sites for one symbol.
search_symbols	Full Text Search symbol names.
list_subclasses	Return the classes that directly inherit from a class.
list_importers	Return the files that import a module.

Resources

Resource	Description
skill://codeparse/SKILL.md	Guided exploration skill for navigating Python codebases.
stats://codeparse/sessions	Per-session markdown table of files with tool calls and tokens saved.

Commands

Command	Description
install	Register the MCP server in installed clients and write the skill.
uninstall	Remove the MCP server and skill.
update	Replace the binary with the latest release and update the skill.
mcp	Start the MCP server.
reload	Drop the index and reprocess the workspace ad-hoc

Benchmark Evaluation

In-Progress

Ten questions found in eval/tasks.json are evaluated against the SQLMesh repo using a baseline agent and an agent with the CodeParse MCP server.

Model	Baseline Avg Tokens	CodeParse Avg Tokens	Token Reduction
Grok 4.7 High Fast	393,530	265,867	-32%
Opus 5.5 Medium	...	...	...

Running the Evaluation

Cursor Default Model: grok 4.7
Claude Defualt Model: opus 5.5 medium

• make eval-smoke cursor will run all 10 questions and evaluate.
• make eval cursor will run all 10 questions on repeat (10 times) and evaluate.