codebase-memory-mcp

name	codebase-memory-mcp
description	用代码图谱索引仓库并查询架构、调用链、依赖和影响面。
metadata	{"homepage":"https://github.com/DeusData/codebase-memory-mcp"}

codebase-memory-mcp

codebase-memory-mcp is a stdio MCP server for code intelligence. It is not a standalone web service. The durable setup is:

1. install the package/binary,
2. register it in the host MCP config,
3. enable auto indexing,
4. let the host start the MCP server for each agent session.

Do not start it as a detached background process just to keep it alive; without an MCP client on stdin/stdout, that process is not useful.

When To Use

Use this skill when the user asks for:

• code graph, repository memory, indexed codebase search, or architecture map,
• "where is this called", "what depends on this", "trace this path",
• cross-file impact analysis before edits,
• large-repo exploration where repeated grep/read cycles would be noisy,
• preserving codebase understanding across future agent sessions.

For exact filenames, literal strings, or quick local search, use rg first. For structural questions, use the MCP tools below after indexing.

Setup / Repair

python -m venv "$env:USERPROFILE\.codebase-memory-mcp-venv"
& "$env:USERPROFILE\.codebase-memory-mcp-venv\Scripts\python.exe" -m pip install --upgrade pip codebase-memory-mcp
& "$env:USERPROFILE\.codebase-memory-mcp-venv\Scripts\codebase-memory-mcp.exe" install
& "$env:USERPROFILE\.codebase-memory-mcp-venv\Scripts\codebase-memory-mcp.exe" config set auto_index true
& "$env:USERPROFILE\.codebase-memory-mcp-venv\Scripts\codebase-memory-mcp.exe" config set auto_index_limit 50000

Verify:

& "$env:USERPROFILE\.codebase-memory-mcp-venv\Scripts\codebase-memory-mcp.exe" config list

Expected config:

auto_index = true
auto_index_limit = 50000

If install says indexes must be rebuilt, confirm only when the listed paths are under the tool's own cache directory, such as ~/.cache/codebase-memory-mcp/.

MCP Workflow

Use MCP tools when available:

1. list_projects to see indexed project names.
2. index_repository with repo_path and a mode:
   • fast for first validation or quick local work,
   • moderate for normal cross-file work,
   • full for deeper architecture/semantic analysis,
   • cross-repo-intelligence when linking already indexed projects.
3. index_status with the project name returned by list_projects.
4. search_graph, trace_path, query_graph, get_code_snippet, get_architecture, or detect_changes depending on the question.

Do not guess the normalized project name. It may be derived from the path, for example D-Toni-code-lotus; read it from index_repository, list_projects, or an index_status hint.

Operating Rules

• Use fast indexing first if the goal is simply to prove the server works.
• Use detect_changes before relying on an older index for active repos.
• Use get_code_snippet or normal file reads before editing any returned path.
• Keep generated persistent artifacts only when the user wants team sharing or reusable code memory in the repo; otherwise use the local cache.
• If the MCP server is unavailable, repair the install once, then fall back to rg, file reads, and the repo's normal tests.