How to use codebase-memory-mcp: install, run and query your code graph

Check whether this tool fits your setup

codebase-memory-mcp is aimed at developers who work with an MCP-capable AI coding agent on a large codebase, engineers who want to see where a feature lives and how calls connect across files, and teams that want to estimate which code and tests a change could affect. It suits people who prefer that source analysis stays on their own machine.

• Indexing and graph queries run on your own machine, and the README says no hosted service or API key is needed.
• Setup assumes some terminal comfort. If you have no MCP-capable agent and no CLI experience, this tool is not a good starting point.
• Smart CI is described as experimental and limited to supported C/C++ test harnesses.

Install codebase-memory-mcp on macOS or Linux

The README's install method uses curl and bash on macOS and Linux. Run the install command below to install the latest release.

• Prerequisites: curl and bash (macOS / Linux).
• The standard native release includes CBM Lens.
• Nix packages and source builds are also available and offer builds without the UI.

Install codebase-memory-mcp on Windows

On Windows the README lists PowerShell as the prerequisite. The Windows install command is not reproduced in the source excerpt this guide is based on, so check the project README for the exact PowerShell command before running anything.

• Prerequisite: PowerShell (Windows).
• Standard releases include CBM Lens for macOS, Linux, and Windows.

Start CBM Lens to explore the graph visually

CBM Lens is the optional visual companion and is in open beta. The MCP server is the core of the project. After installation you can start CBM Lens with the command below, which opens it in your browser on port 9749.

CBM Lens includes a source reader, architecture maps, the Galaxy 3D graph, coverage diagnostics, architecture decisions, and local code chat. Its local chat can use a model that runs on your device once you enable it.

• Folders appear as nested platforms, and files and symbols as selectable blocks with arrows for indexed relationships.
• Explore shows files in an IDE-style workspace with syntax-highlighted source beside file impact and symbol callers and calls.

Connect the MCP server to your AI coding agent

The MCP server is the core of the project. It exposes 17 MCP tools that let an agent ask structural questions, such as what calls a function or what depends on a file, instead of reading many files. The installer supports 45 automatic or conditional client integrations, including client configuration, skills, agent profiles, and applicable hooks.

Support varies by client, so check the integration matrix in the README for what your specific agent receives.

• The installer writes agent configuration for the clients you select.
• Indexes persist between sessions, and eligible active MCP sessions can keep them synchronized.
• You can also query the graph directly from the CLI.

Use the graph for everyday code questions

Ask focused questions rather than broad ones. The project says targeted graph queries can reduce repeated file reads and token use, and the README links to scoped measurements in its Performance section. The tool also supports symbol, text and semantic search, callers, dependencies, source snippets and file outlines.

Documentation is indexed too: document sections, PDF text pages, architecture decisions and supported references can be connected to code, so you can find an explanation next to the symbol it describes.

• Find where a feature lives and how calls connect across files.
• Trace supported relationships through Docker, Kubernetes, Helm, Kustomize, Terraform/OpenTofu, Terragrunt, Ansible, scripts, Make, and CI configuration.
• Inspect affected symbols and compare indexed snapshots before running a change.

Estimate what a change affects with Smart CI

Smart CI is experimental. It combines static relationships and per-test coverage for supported C/C++ test harnesses and falls back to broader runs when evidence is insufficient. Do not rely on it for languages beyond the supported C/C++ harnesses.

• Treat it as an estimate and check results for your project.
• Fallback to broader test runs is intentional when evidence is weak.