Deepcontext MCP

🚀 Smarter Context for Large Codebases

DeepContext is an MCP server that adds symbol-aware semantic search to Codex CLI, Claude Code, and other agents. It provides these agents with a more precise context of even the largest codebases. Currently, it supports Typescript and Python.

🚀 Quick Start

1. Visit the Wildcard DeepContext page.
2. Click "Generate API Key".
3. Copy your API key.
4. Paste the installation command for your MCP client.
5. Type index this codebase to index the current directory.

Claude Code:

claude mcp add deepcontext \
  -e WILDCARD_API_KEY=your-wildcard-api-key \
  -- npx @wildcard-ai/deepcontext@latest

Codex:

# Add to ~/.codex/config.toml
[mcp_servers.deepcontext]
command = "npx"
args = ["-y", "@wildcard-ai/deepcontext@latest"]
env = { "WILDCARD_API_KEY" = "your-wildcard-api-key" }

🎥 Demo

Check out the demo

✨ Features

Most coding agents use grep-based search that matches exact text. These searches miss semantically related code and fill context windows with irrelevant results. Large codebases amplify this problem, where text search returns hundreds of matches that quickly overwhelm conversation capacity. This leads to slow completions, more hallucinations, and lower success rates.

DeepContext provides agents with intelligent search that preserves context windows by finding only relevant code chunks.

• Semantic accuracy: Matches code by meaning and relationships rather than text patterns, finding related functions across files that keyword search misses.
• Reduced token usage: Returns precise code chunks instead of every file containing your search terms, preserving conversation context windows and reducing costs.
• Search speed: Searches code immediately through pre-indexed data for instant file discovery.

🛠️ MCP Tools

index_codebase

Creates a searchable index of your codebase for semantic search.

search_codebase

Finds relevant code using natural language or keyword queries.

get_indexing_status

Shows indexing status and file counts for your codebases.

clear_index

Removes all indexed data for a codebase.

🔧 Technical Details

MCP Integration Flow

• The coding agent communicates with DeepContext through the Model Context Protocol.
• The MCP server receives requests, validates parameters, and routes them to appropriate core components.
• For long-running operations like indexing, it spawns detached background processes to prevent timeouts. Background workers handle large codebases without blocking the MCP channel.

AST-Based Parsing

• Tree-sitter parsers analyze source code to build Abstract Syntax Trees. It uses Python, TypeScript, and JavaScript language grammars for accurate parsing and identifies semantic nodes for functions, classes, interfaces, and modules.
• Symbol extraction identifies functions, classes, interfaces, types, variables, and constants. Scope analysis determines local, exported, and global visibility, and parameter and return type extraction is done for function signatures.
• Import/export analysis maps module dependencies and cross-file relationships. It creates chunks at semantic boundaries rather than arbitrary line or token splits and handles large files through range-based parsing with overlapping windows.

Hybrid Search with Reranking

• Search operates in three stages. Hybrid search combines vector similarity and BM25 full-text search, and the Jina reranker-v2 is used for final relevance optimization.
• Vector similarity finds semantically related code using embeddings. Jina text embeddings generate 1024 - dimension vectors for code chunks.
• BM25 performs traditional keyword matching for exact terms. Full-text indexing enables precise identifier and comment matching.
• Results are fused using configurable weights and then reordered by the Jina reranker.

Incremental Indexing

• It uses file modification times and content hashes to track changes. SHA - 256 hashing detects content modifications at the byte level.
• Only files with different hashes are reprocessed during reindexing, avoiding unnecessary parsing and embedding generation for unchanged files.

Content Filtering

• Files are scored based on extension patterns, path components, and content analysis. Language detection and file type classification are used for processing decisions.
• Test files, generated code, minified files, and build outputs are excluded during indexing. Pattern matching against common test frameworks and build tool outputs is done.
• Documentation and configuration files are filtered to focus on source code.

🏠 Self Hosting

Self - hosting requires code modifications to integrate directly with vector storage and embedding providers, as the current implementation uses the Wildcard API backend.

Prerequisites

• Node.js 20+ for ES module support and performance optimizations.
• Turbopuffer API key for vector storage and hybrid search operations.
• Jina AI API key for text embeddings and reranking services.

Setup

git clone https://github.com/Wildcard-Official/deepcontext-mcp.git
cd deepcontext
npm install
npm run build

Integration

claude mcp add deepcontext-local \
  -e TURBOPUFFER_API_KEY=your-turbopuffer-key \
  -e JINA_API_KEY=your-jina-key \
  -- node /path/to/deepcontext/dist/standalone-mcp-integration.js

🤝 Contributing

Thanks for your interest! We’re currently not accepting external contributions as we’re an early - stage startup focused on rapid iteration. We may open things up in the future — feel free to ⭐ the repo to stay in the loop.

📄 License

Licensed under the Apache License.