# DeusData codebase-memory-mcp Dependencies: Complete Vendored Library Reference

> Explore the vendored dependencies for DeusData codebase-memory-mcp. Discover a self-contained binary with no external package managers or system libraries needed for installation.

- Repository: [Martin Vogel/codebase-memory-mcp](https://github.com/DeusData/codebase-memory-mcp)
- Tags: api-reference
- Published: 2026-07-19

---

**DeusData codebase-memory-mcp ships as a fully static binary with all runtime dependencies vendored, requiring no external package managers, system libraries, or internet connectivity at install time.**

The `DeusData/codebase-memory-mcp` repository provides a single-binary MCP server that embeds every library it needs for parsing, indexing, and semantic search. Because **DeusData codebase-memory-mcp dependencies** are compiled directly into the executable at build time, end users never execute `npm install`, `pip install`, or `apt-get` to deploy the tool.

## Core Parsing and Storage Dependencies

The foundation of the indexing pipeline relies on Tree-sitter for AST generation and SQLite for persistent storage.

### Tree-sitter Runtime and Language Grammars

At `internal/cbm/vendored/ts_runtime`, the project embeds the **Tree-sitter C runtime**, which serves as the core parser generator producing ASTs for 158 supported languages. The specific language grammars—160 pre-generated parsers including `tree-sitter-python` and `tree-sitter-go`—reside in `internal/cbm/vendored/grammars`. These components are compiled into the binary, eliminating runtime linking against external parser libraries.

### SQLite Graph Database

Persistent graph storage—including nodes, edges, and FTS5 full-text search—depends on the public-domain SQLite engine located at `vendored/sqlite3`. The database driver is statically linked, allowing the binary to create and query `*.db` files without requiring a system SQLite installation.

## Indexing Pipeline Utilities

High-throughput code analysis requires specialized memory management, compression, and data processing libraries.

### Memory Allocation and Compression

For high-performance indexing, the project uses **mimalloc** (`vendored/mimalloc`), a fast, low-fragmentation memory allocator that reduces heap overhead during batch processing. Source code compression utilizes **LZ4** (`internal/cbm/vendored/lz4`) for in-memory compression during indexing, while **Zstandard** (`internal/cbm/vendored/zstd`) handles final on-disk compression of SQLite graph databases, producing `*.zst` artifacts.

### Hashing and Deduplication

Incremental indexing and deduplication rely on multiple high-speed hash implementations. **xxHash** (`vendored/xxhash`) provides rapid hashing for file content comparison. **Verstable** (`internal/cbm/vendored/verstable`) generates versioned stable hashes for graph nodes to support incremental updates. **wyhash** (`internal/cbm/vendored/wyhash`) offers low-overhead hashing for internal data structures.

### Text Processing and Parsing

JSON parsing for the MCP JSON-RPC protocol uses **yyjson** (`vendored/yyjson`), while POSIX-compatible regular expressions for structural searches are handled by **TRE** (`vendored/tre`). For C/C++ header analysis, the **simplecpp** preprocessor (`internal/cbm/vendored/simplecpp`) parses macro definitions and include directives before Tree-sitter AST extraction.

## Semantic Search and ML Components

Beyond syntax analysis, the tool provides neural code search through an embedded machine learning model.

### Nomic Embeddings Model

The semantic search functionality utilizes the **nomic-embed-code** model, a 40K-token, 768-dimensional embedding system stored in `vendored/nomic`. Licensed under Apache 2.0, this model enables the `cli semantic_query` command to perform vector similarity searches without calling external APIs. When you execute:

```bash
codebase-memory-mcp cli semantic_query '{"query":"authentication"}'

```

The binary references the bundled embeddings stored in `vendored/nomic` to compute vector similarities against indexed code.

## Optional Frontend Dependencies

The core binary operates independently of any JavaScript runtime. However, the optional 3-D visualization interface requires additional frontend libraries.

### React and Three.js UI Stack

When building the `-ui` variant, the project bundles a React 18 application defined in [`graph-ui/package.json`](https://github.com/DeusData/codebase-memory-mcp/blob/main/graph-ui/package.json). This interface uses **Three.js** and **@react-three/fiber** for 3-D graph rendering, **radix-ui** for accessible UI components, **lucide-react** for iconography, and **Tailwind CSS** for styling. These dependencies are compiled into static assets served at `localhost:9749` when running:

```bash
codebase-memory-mcp --ui=true --port=9749

```

Note that the standard binary does not include these web assets; they are only present in builds specifically marked with the UI variant.

## Dependency Verification and Licensing

All vendored code is inventoried in [`THIRD_PARTY.md`](https://github.com/DeusData/codebase-memory-mcp/blob/main/THIRD_PARTY.md) at the repository root. This document provides the complete licensing breakdown for every embedded library, from BSD-2-Clause hash implementations to MIT-licensed JSON parsers.

To inspect the exact dependency versions compiled into your binary:

```bash

# Display version information confirming embedded runtime presence

codebase-memory-mcp --version

# View the complete third-party attribution list

cat THIRD_PARTY.md

```

The [`pkg/npm/package.json`](https://github.com/DeusData/codebase-memory-mcp/blob/main/pkg/npm/package.json) file contains NPM packaging metadata showing the binary entry point, though this is used only for distribution wrapping and does not affect the binary's internal vendored dependencies.

## Summary

- **DeusData codebase-memory-mcp dependencies** are entirely vendored and statically compiled—no runtime package installation is required.
- Core functionality relies on Tree-sitter (`internal/cbm/vendored/ts_runtime`), SQLite (`vendored/sqlite3`), and mimalloc (`vendored/mimalloc`).
- Compression, hashing, and JSON processing use LZ4, Zstandard, xxHash, and yyjson, all embedded in `vendored/` and `internal/cbm/vendored/` directories.
- Semantic search operates via the bundled **nomic-embed-code** model in `vendored/nomic`, enabling offline vector queries.
- The optional 3-D UI ([`graph-ui/package.json`](https://github.com/DeusData/codebase-memory-mcp/blob/main/graph-ui/package.json)) requires React and Three.js, but only when explicitly building the UI-enabled binary variant.
- Authoritative licensing and dependency information resides in [`THIRD_PARTY.md`](https://github.com/DeusData/codebase-memory-mcp/blob/main/THIRD_PARTY.md).

## Frequently Asked Questions

### Does codebase-memory-mcp require npm, pip, or system package managers to run?

No. The binary embeds all required C libraries, Tree-sitter grammars, and the SQLite engine. You only need to download the archive for your OS and architecture. The optional UI variant includes pre-compiled JavaScript assets and does not require `npm install` at runtime.

### Where are third-party libraries located in the source repository?

Vendored dependencies reside in two primary locations: `vendored/` for top-level components like SQLite and xxHash, and `internal/cbm/vendored/` for internal libraries like Tree-sitter grammars, LZ4, Zstandard, and simplecpp. The complete inventory is documented in [`THIRD_PARTY.md`](https://github.com/DeusData/codebase-memory-mcp/blob/main/THIRD_PARTY.md).

### What embedding model powers the semantic search feature?

The semantic search tool uses the **nomic-embed-code** model, a 40K-token, 768-dimensional embedding system stored in `vendored/nomic`. This Apache 2.0-licensed model is bundled directly into the binary, allowing offline vector similarity searches without external API calls.

### Is internet connectivity required after installation?

No. Because all dependencies—including the Tree-sitter grammars, SQLite database engine, and Nomic embedding model—are vendored into the static binary, the tool operates entirely offline after the initial binary download.