A code search tool for AI coding agents that reduces token consumption by using a static embedding model on CPU, available as an MCP server, CLI, or Python library with no required external services.
Project overview
It addresses the token cost of codebase exploration by running locally on CPU without API keys or GPU, reporting approximately 98% fewer tokens than combined grep and file-read workflows.
Project type
MCP
Deployment
Refer to project documentation
License
MIT
Best for
Developers and AI engineers who need to provide coding agents with a tool to search local or remote repositories using CPU-only infrastructure.
Key capabilities
Accepts natural language queries, code queries, local file paths, or remote git URLs to search and return relevant code snippets along with matched file paths and line numbers.
Runs as an MCP server, allowing coding agents to execute searches directly through native tool calls.
Provides a standalone CLI for searching codebases within shell scripts or outside of an active MCP session.
Locates code chunks semantically related to the code located at a specified file path and line number.
Tracks and displays an estimate of the tokens saved across searches compared to reading full files.
Reads .gitignore and .sembleignore files to determine which files to exclude or include during indexing.
Limitations and risks
The tool relies on a static embedding model and does not perform a transformer forward pass at query time.
Telemetry practices are not documented in the available facts.
Getting started
Install via uv tool install semble. Run semble install to start the interactive installer, which configures agents automatically. Coding expertise is not required.