codebase-time-machine 0.1.2

Understand why code exists, not just what it does - an LLM-agnostic tool server for code history analysis

CTM MCP Server

The MCP (Model Context Protocol) server that powers Codebase Time Machine. Use this with Claude Code, Claude Desktop, or any MCP-compatible client.

Quick Start

Option A: Install from PyPI (Recommended for Users)

pip install codebase-time-machine

# Or with pipx (isolated environment)
pipx install codebase-time-machine

Note: Install in your base Python environment (not a .venv) so Claude Code and Claude Desktop can access it.

Claude Code config - create a .mcp.json file in your project root:

    {
  "mcpServers": {
    "ctm": {
      "command": "python",
      "args": ["-m", "ctm_mcp_server.stdio_server"],
      "env": {
        "GITHUB_TOKEN": "your_token_here"
      }
    }
  }
}

Then restart Claude Code. It will automatically detect and use the MCP server.

Option B: Run from Source (For Development)

git clone https://github.com/BurakKTopal/codebase-time-machine.git
cd codebase-time-machine
uv sync

The repo includes a .mcp.json that configures Claude Code automatically:

{
  "mcpServers": {
    "ctm": {
      "command": "uv",
      "args": ["run", "ctm_server.py"]
    }
  }
}

Just open the project in Claude Code and the MCP server is available.

Claude Desktop Setup

Add to your Claude Desktop config:

Config locations:

• macOS: ~/Library/Application Support/Claude/claude_desktop_config.json
• Windows: %APPDATA%\Claude\claude_desktop_config.json
• Linux: ~/.config/Claude/claude_desktop_config.json

If installed via pip:

{
  "mcpServers": {
    "ctm": {
      "command": "python",
      "args": ["-m", "ctm_mcp_server.stdio_server"],
      "env": {
        "GITHUB_TOKEN": "your_token_here"
      }
    }
  }
}

If running from source:

{
  "mcpServers": {
    "ctm": {
      "command": "uv",
      "args": ["--directory", "/path/to/codebase-time-machine", "run", "ctm_server.py"],
      "env": {
        "GITHUB_TOKEN": "your_token_here"
      }
    }
  }
}

GitHub Token (Optional)

The GITHUB_TOKEN is optional for public repositories but recommended for:

• Private repository access
• Higher API rate limits (5000/hour vs 60/hour)

Get one at: https://github.com/settings/tokens (needs repo scope for private repos)

VS Code Extension

Prefer a UI? The VS Code Extension provides right-click code investigation without needing Claude Code or Desktop. It uses this MCP server under the hood to provide tools to the AI agent. Includes follow-up questions and multi-provider support (Anthropic, OpenAI, Gemini).

Tools Reference

CTM provides 32 tools for code investigation:

Primary Tools (Start Here)

Tool	Speed	Description
get_line_context	2-4s	Flagship tool - Why does this line exist? Returns blame, commit, PR, issues, discussions
get_local_line_context	2-4s	Same as above, for local repos (auto-bridges to GitHub if remote exists)
get_github_file	<1s	Get file contents
explain_file	3-5s	File overview: purpose, symbols, contributors, recent changes

GitHub Tools

Tool	Description
get_github_file_history	Commits that modified a file
get_github_file_symbols	Extract functions/classes from a file
get_github_commit	Get commit details
get_github_commits_batch	Fetch multiple commits at once (faster)
get_github_repo	Repository information
get_github_branches	List branches
list_github_tree	Browse repository file structure
get_pr	Pull request details with comments and reviews
get_issue	Issue details with comments
search_prs_for_commit	Find PRs containing a commit
search_github_code	Search code in repository
search_github_commits	Search commit messages
pickaxe_search_github	Find when code was added/removed
trace_github_symbol_history	Track function/class evolution
get_code_owners	Top contributors for a file
get_change_coupling	Files that frequently change together
get_activity_summary	Repository/path activity overview

Local Git Tools

Tool	Description
get_repo_info	Repository metadata
list_branches	List branches with last commit
get_commit	Commit details
get_commit_diff	Detailed diff for a commit
trace_file_history	Complete file change history
get_file_at_commit	File contents at specific commit
pickaxe_search	Find when code was added/removed
explain_commit	Analyze commit intent (bugfix, feature, refactor)
blame_with_context	Git blame with commit metadata
get_file_symbols	Extract functions/classes from local file
trace_symbol_history	Track symbol changes in local repo

---

Design Philosophy

Powerful Flagship Tools

Instead of simple wrappers around git commands, CTM tools are designed to aggregate multiple data sources in a single call:

get_line_context / get_local_line_context - The flagship tool that:

• Runs git blame to find who last modified the code
• Automatically runs per-line pickaxe search to find the true origin of each line (when code was first introduced)
• Groups origins by commit SHA for efficient multi-line selections
• Detects if code was introduced as a comment vs active code
• Fetches the associated PR (if the commit came from a PR)
• Extracts linked issues from PR/commit messages
• Detects patterns (commented code with active alternatives, TODOs, stale fixes)
• Provides quick answers when patterns are clear (e.g., "This is unused alternative code")
• Shows nearby context to detect active implementations below commented code
• Calculates confidence scores with specific signals for investigation quality
• Returns everything in one structured response with code_sections containing grouped origins

This is the difference between "a git wrapper" and "a code investigation tool".

Intelligent Caching

CTM recognizes that git data has different mutability characteristics:

Data Type	TTL	Reason
Commits	Forever	Commits are immutable
File contents at commit	Forever	Immutable (content-addressed)
Git trees	Forever	Immutable
Repository metadata	24 hours	Rarely changes
PRs and issues	1 hour	May get new comments
Search results	30 minutes	Index updates

Result: Repeat investigations are nearly instant, and GitHub API rate limits are preserved.

Batch Operations

Tools like get_github_commits_batch fetch multiple items in parallel:

• First call: 5 commits → 5 API calls (cached)
• Second call: Same 5 commits → 0 API calls (cache hits)
• Mixed call: 3 cached + 2 new → 2 API calls

This makes follow-up investigations significantly faster.

---

Configuration

Environment Variables

Variable	Default	Description
GITHUB_TOKEN	(none)	GitHub personal access token
CTM_CACHE_PATH	~/.ctm/cache.db	SQLite cache location

Caching

CTM uses SQLite for persistent caching at ~/.ctm/cache.db.

TTL strategy:

• Commits, git trees: never expire (immutable)
• File contents at specific commits: never expire (immutable)
• Repository metadata: 24 hours
• PRs and issues: 1 hour
• Search results: 30 minutes

Agent Guide

For optimal tool usage, add CLAUDE.md to your project. It teaches LLMs:

• Which tools are fast vs slow
• Tool selection for different questions
• Caching strategies and batch operations

Troubleshooting

"Module not found"

python -c "import ctm_mcp_server; print('OK')"

"Module not found" in Claude Desktop (Python version mismatch)

If the package is installed but Claude Desktop can't find it, the python command may resolve to a different Python version than the one used to install the package.

Solution: Specify the full path to the Python executable in your config:

{
  "mcpServers": {
    "ctm": {
      "command": "path/to/python.exe",
      "args": ["-m", "ctm_mcp_server.stdio_server"],
      "env": {
        "GITHUB_TOKEN": "your_token_here"
      }
    }
  }
}

To find your Python path, run where python (Windows) or which python (macOS/Linux), and use the path where you installed the package.

"Rate limit exceeded"

Add a GITHUB_TOKEN for 5000 requests/hour instead of 60.

"Repository not found"

For private repos, ensure your token has repo scope.

License

AGPL-3.0