An MCP server that gives AI coding tools IDE-like navigation and precise, token-efficient refactors.
Understand Why Features Aren't Shipping Faster
In 15 minutes, see exactly where engineering time goes - using only Git metadata, never your source code.
- Indexes a workspace for fast symbol search and navigation
- Exposes safe, structured edits (
find_references,move_symbol,change_signature) - Supports Python, JavaScript, TypeScript, Go, HTML, CSS
- Rename a symbol across 10 files without collecting context or crafting diffs for each file. Rough savings: each file still needs context for safe patching; even 20–40 lines per file (~400–800 tokens) is ~4k–8k tokens avoided at 10 files.
- Change a function signature and update call sites with one tool call instead of manual diffs for each call site.
- Move a symbol between files and update imports without loading both files to build the patch.
The MCP client will launch the server on demand using the command below.
Codex CLI:
codex mcp add llm-ide -- /path/to/repo/.venv/bin/python -m refactor_mcp.serverClaude Code:
claude mcp add llm-ide -- /path/to/repo/.venv/bin/python -m refactor_mcp.serverIf your repo uses CLAUDE.md, add a short note to encourage token-saving tool use:
## MCP (LLM-IDE)
Use the LLM-IDE MCP tools proactively to reduce token usage.
Prefer structured operations (find references, rename, move, signature changes)
over manual diff editing. Fetch only needed context (symbols/ranges) instead of
reading whole files.cd /path/to/refactor_mcp
python3 -m venv .venv
source .venv/bin/activate
pip install -U pip
pip install -e '.[dev]'docker build -t llm-ide .
docker run -p 8000:8000 -v /path/to/your/repo:/workspace llm-ideREFACTOR_MCP_DISABLE_WATCHER=1disables the file watcher.REFACTOR_MCP_AUTO_INDEX=0disables auto indexing oninitialize_workspace.REFACTOR_MCP_FORCE_REINDEX=1forces a full re-index every time.