CodeIndex.Mcp 1.0.7

CodeIndex.Mcp

MCP server for semantic code indexing with Qdrant and llama.cpp.

Changelog

1.0.7

• Added runtime controls for expensive background work:
  • New --disable-rerank flag disables the local reranker service and removes rerank from advertised search modes
  • New --filewatcher-interval-seconds option controls the delay between background FileWatcher passes
  • get_index no longer requires the reranker service unless search_mode=rerank is requested
  • MCP search_mode schema default now matches runtime default: hybrid

1.0.6

• Breaking Change: Replaced adaptive HTTP concurrency with static --workers parameter:
  • New CLI option --workers (default: 2, min: 1, max: 32)
  • Controls both HTTP request concurrency (tokenize + embed) and llama embedding workers
  • Removed adaptive HTTP limiter and health-based dynamic adjustment
  • Simplified progress logging: removed tok/emb/lim/health counters, kept only ok/err

1.0.5

• Optimized external llama server throughput for multi-GPU setups:
  • Increased HTTP parallelism from CPU *2 to CPU *4 for external llama mode
  • Increased backpressure queue capacity from 1024 to 2048
  • Added explicit HttpClient connection pooling (64 connections,5min lifetime)

1.0.4

• Merged c++ and c files into the same category
• Deleted fusion field from response of get_index method

1.0.3

• Added centralized TextNormalization utility class for consistent text handling
• Consolidated line ending normalization (CRLF→LF) and path separator normalization (\→/)
• Upgraded file path normalization, taking into account quotation marks returned by git commands
• Improved code maintainability by removing duplicate normalization methods

Features

• Semantic Code Search: Index and search code using vector embeddings
• Git Repository Support: Automatic file discovery via git ls-files
• Tree-sitter Chunking: Intelligent code chunking based on AST
• Multiple Embedding Backends:
  • Local llama.cpp (GPU accelerated via CUDA/Vulkan/SYCL)
  • External OpenAI-compatible APIs (OpenRouter, etc.)
  • Remote llama.cpp server
• Hybrid Search: Dense + sparse (BM25) vectors with RRF/DBSF fusion
• Reranking: Optional local llama.cpp reranker for improved relevance
• Distributed Indexing: Shard support for parallel one-shot indexing

Installation

dotnet tool install -g CodeIndex.Mcp

Usage

MCP Mode (for AI assistants)

code-index-mcp

The server communicates via JSON-RPC over stdio, following the MCP protocol.

One-shot Mode (single indexing run)

code-index-mcp --repo C:\path\to\repo --one-shot

Show Version

code-index-mcp --version

Command Line Options

Core Options

Local llama.cpp Options

Used for local embedding (when --external-embed-api-key is not set) and for reranking unless --disable-rerank is set.

External Embedding API Options

When --external-embed-api-key is set, embedding is performed via external API. Reranking still uses local llama.cpp unless --disable-rerank is set.

External Types

• service: OpenRouter/OpenAI-compatible payload. Includes model field, and provider field per OpenRouter rules.
• llama: llama-compatible payload. Uses --llama-embed-model for model name, no provider field. For remote llama.cpp servers.

Distributed One-shot Options

For parallel indexing across multiple processes/machines. Only valid with --one-shot.

Files are distributed by: ordered_file_index % shard_count == shard_index

Requirements

• Qdrant: Running instance at http://localhost:6333 (or custom URL via --qdrant-url)
• llama.cpp (for local embedding/reranking):
  • C:\llama\cuda\llama-server.exe for NVIDIA
  • C:\llama\vulkan\llama-server.exe for AMD
  • C:\llama\sycl\llama-server.exe for Intel
  • C:\llama\models\jina-code-embeddings-0.5b-Q8_0.gguf (embedding)
  • C:\llama\models\jina-reranker-v3-Q8_0.gguf (reranking, unless disabled)

MCP Tools

The server provides these MCP tools:

set_index

Index or update repository code.

{"name": "set_index", "arguments": {"reindex": false}}

• reindex (bool, default: false): Full reindex - deletes and recreates both collections

get_index

Semantic code search.

{"name": "get_index", "arguments": {"query": "search query", "count": 32, "search_mode": "rerank"}}

• query (string, required): Search query
• count (int, default: 32, range: 16-64): Number of results
• path_prefix (string, default: empty): Filter by path prefix
• lang (string, default: empty): Filter by programming language
• search_mode (string, default: hybrid): dense, sparse, hybrid, or rerank

recovery

Cleanup inconsistent state after errors.

{"name": "recovery", "arguments": {}}

Removes stale entries from collections based on current repository state.

Examples

Basic MCP Mode

code-index-mcp

One-shot Indexing

code-index-mcp --repo C:\projects\myapp --one-shot

External Embedding API (OpenRouter)

code-index-mcp --repo C:\projects\myapp `
  --external-embed-api-key sk-xxx `
  --external-embed-api-url https://openrouter.ai/api/v1 `
  --external-type service `
  --service-embed-model openai/text-embedding-3-small

External llama.cpp Server

code-index-mcp --repo C:\projects\myapp `
  --external-embed-api-key any `
  --external-embed-api-url http://192.168.1.100:8080 `
  --external-type llama

Force NVIDIA GPU

code-index-mcp --repo C:\projects\myapp --vendor nvidia

Distributed Indexing (2 shards)

# Terminal 1 - Shard 0
code-index-mcp --repo C:\projects\myapp --one-shot --scan-shard-index 0 --scan-shard-count 2

# Terminal 2 - Shard 1
code-index-mcp --repo C:\projects\myapp --one-shot --scan-shard-index 1 --scan-shard-count 2

License

MIT