pow-mcp-rag-new

A local RAG (Retrieval-Augmented Generation) MCP server for project documentation and code.

Overview

This project provides a Model Context Protocol (MCP) server that indexes project documentation (specs, headers, source files, PDFs, configs) into a local vector database (ChromaDB) and exposes semantic search tools to Kiro or any MCP-compatible client.

Supports multiple tech stacks: C/C++, Python, Go, C#, Node.js/TypeScript.

Distribution name: pow-rag-mcp (unrelated to rag-mcp or rag-mcp-server packages).

Minimum Python version: 3.11 (matches requires-python = ">=3.11" in pyproject.toml).

Three deployment modes:

• Docker (Phase 1) — zero local Python needed; server runs in a container

• PyPI (recommended for new users) (Phase 2a) — install via uvx --from, uv tool install, or pip install from PyPI; no repo checkout or local index needed

• Local PyPI + uvx (Phase 2b) — install rag-mcp via uvx from a local package index; no persistent venv, easiest to keep updated. Stepping stone toward a hosted index (S3/CodeArtifact).

• pip install (Phase 2c) — install rag-mcp directly into a Python environment

Related MCP server: ragi

License

This project is licensed under the Apache 2.0 License — see the LICENSE file for details.

Quick Start

PyPI (recommended for new users)

Install from PyPI using your preferred method:

Option 1: One-off execution with uvx

uvx --from pow-rag-mcp rag-mcp serve

Option 2: Persistent install with uv tool install

uv tool install pow-rag-mcp

Option 3: Traditional pip

pip install pow-rag-mcp

All three methods fetch pow-rag-mcp from PyPI directly — no repository checkout or local package index is required.

See doc/PIP_INSTALL_GUIDE.md for full details (config seeding, bundled docs, upgrades, and migrating to a hosted index).

Local PyPI + uvx (recommended for development/publishing workflows)

cd <your-checkout>/pow-mcp-rag-new
setup-pypi.bat

This builds the wheel, publishes it to a local pypiserver index (packages/), installs it as a persistent uv tool (a stable exe at ~/.local/bin/rag-mcp.exe — resolved once, not on every launch), and wires up ~/.kiro/settings/mcp.json + .vscode/mcp.json to launch it directly. No Docker, no persistent venv to manage. See TROUBLESHOOTING.md for why this is preferred over plain uvx --from on Windows with this package's large dependency tree.

Full guide: doc/PIP_INSTALL_GUIDE.md (keep the local index running, rebuild/republish after code changes, migrating to S3 later).

Docker (alternative)

cd <your-checkout>/pow-mcp-rag-new
docker build -t rag-mcp-new-pip:latest .

# Index your projects (set SRC to your repos folder)
$SRC = "C:/Users/you/GIT"   # PowerShell
docker run --rm -v "${SRC}:/projects:ro" -v rag-mcp-new-pip-data:/app/data rag-mcp-new-pip:latest python indexer.py

Then add to ~/.kiro/settings/mcp.json:

{
  "mcpServers": {
    "rag-mcp": {
      "command": "docker",
      "args": [
        "run", "-i", "--rm",
        "-v", "C:/Users/you/GIT:/projects:ro",
        "-v", "rag-mcp-new-pip-data:/app/data",
        "rag-mcp-new-pip:latest",
        "python", "server.py", "--no-reindex"
      ],
      "disabled": false,
      "autoApprove": [
        "search_docs", "search_specs", "search_code", "search_logs",
        "list_projects", "list_files", "get_document", "get_project_summary",
        "find_function", "find_variable", "search_hex_pattern", "compare_projects",
        "add_project", "add_file", "add_folder", "add_pattern", "index_log_file"
      ]
    }
  }
}

Restart Kiro — the RAG is ready. Full guide: doc/DOCKER_GUIDE.md (project management, HTTP server, docker run CLI reference, setup-docker.bat automation).

pip install (hosted index, once available)

pip install torch --index-url https://download.pytorch.org/whl/cpu
pip install rag-mcp --extra-index-url https://<your-s3-bucket>.s3.<region>.amazonaws.com/
rag-mcp config   # seeds config on first run, shows resolved paths
rag-mcp index
rag-mcp serve

Full guide: doc/PIP_INSTALL_GUIDE.md

Documentation

Reading this from pip show / a package-index page instead of the repo? This README's links are relative paths into the pow-mcp-rag-new repo (GitHub/clone) and won't resolve from an installed package alone. Three docs travel with the install and are available offline via rag-mcp docs <name> (see table below); the rest require the repo checkout.

Run rag-mcp docs with no arguments to list the bundled docs from any install (pip, uvx, or uv tool install) without needing the repo.

What gets indexed

File types are configured in config.yaml under index_extensions. By default: C/C++, Python, React/JS/TS, C#, Go, Kotlin, Markdown, PDF, text. Directories in excluded_dirs (build, node_modules, .git, ...) are skipped. See doc/DOCKER_GUIDE.md for details on configuring projects and adding new sources.

MCP Tools

Once configured, 21 MCP tools are available for searching, browsing, and managing your indexed projects — see doc/TOOLS_GUIDE.md for the full list with examples (also available offline: rag-mcp docs tools).

Key tools:

• search_docs / search_specs / search_code — semantic search, scoped by file type

• search_hex_pattern, find_function, find_variable — exact text matching for codes/symbols

• search_logs, index_log_file — structured log search and on-demand indexing

• add_project, add_pattern, remove_project — index management from Kiro chat

Log Indexing

Structured log indexing with severity filtering, time-window search, and hex error-code matching. The pipeline is fully generic — driven by YAML pattern configs, so any log format can be supported without code changes. See doc/DOCKER_GUIDE.md and doc/LOG_INDEXING_GUIDE.md (also: rag-mcp docs log-indexing).

Need Help?

See doc/TROUBLESHOOTING.md for common issues, including:

• MCP server hangs on startup

• Stale search results after re-indexing

• entrypoint.sh exec errors from CRLF line endings

• --project <NAME> silently doing nothing

• Concurrent query errors