Synapse Code MCP

What Eltortilla1/synapse-code-mcp MCP server does

Eltortilla1/synapse-code-mcp MCP server connects an AI assistant to a repository on the local machine. Its tools expose repository structure and source-code context through the Model Context Protocol, so an agent can inspect a project without requiring files to be copied into a prompt.

The server focuses on structural and textual code understanding. It can return a compressed map of exported functions, classes, interfaces, types, enums, and top-level constants; provide a file together with locally related dependencies; search source files; show a repository tree; and report files changed since a Git reference. It does not provide conceptual search through embeddings or a vector database.

How it works

The server reads the configured project directory and uses filesystem and Git information to answer MCP tool calls. The project tree honors .gitignore rules. Code searches support plain text and regular expressions and return matching paths and line numbers. Git changes are grouped by status and can include a unified diff.

For TypeScript and JavaScript, analysis uses the TypeScript compiler API through ts-morph. Dependency traversal follows imports and requires, including relative imports and configured TypeScript path aliases when a root tsconfig.json is available. A depth limit controls how far dependency traversal proceeds. Python, Go, and Rust receive regex-based extraction for names, while dependency traversal is not available for those languages. Other languages can still be read and searched.

The project index is intended as an initial orientation step for unfamiliar repositories. It omits implementation bodies and reports signatures or names, reducing the amount of source content returned. Large repositories can be narrowed with a file glob.

Setup and configuration

Install the published package globally with npm, or run it through npx. The package requires Node.js 18 or newer. When configuring a client, start the server with the repository root supplied through --root; use an absolute path for predictable client behavior.

Claude Desktop, Claude Code, Cursor, and Windsurf are documented client options. Each configuration starts the package with an MCP server entry and points it at the repository to inspect. Multiple instances can be configured for separate project roots.

The available command-line settings include the project root and options for controlling server behavior; the tool parameters provide the main per-request configuration. Relevant request controls include file patterns, maximum tree depth, hidden-file display, dependency depth, output format, and search result limits.

Tools and capabilities

• get_project_tree returns a structured repository tree while respecting ignore rules.
• get_project_index returns a compact semantic map and supports glob scoping, inclusion of non-exported symbols, and Markdown or JSON output.
• get_semantic_context returns file content with local dependency context, with an outline-only mode for signatures and configurable import depth.
• search_codebase searches text or regular expressions and reports file paths and line numbers. It uses ripgrep when available and otherwise falls back to a Node.js scanner.
• get_changed_files groups added, modified, deleted, or renamed files since a Git ref and can include line counts and the full diff.

Limitations and notes

Eltortilla1/synapse-code-mcp MCP server is an early-stage project, according to the README. Its dependency graph features are limited to TypeScript and JavaScript. For Python, Go, and Rust, the index and outline provide names rather than full TypeScript-style signatures, and other languages do not receive the same structural extraction.

The project index grows with the number of symbols, so monorepos and large repositories should be scoped with file_pattern. The server is local and does not require an API key, embedding provider, or vector database, but it also cannot perform natural-language conceptual retrieval across code. Choose it for structure, text, dependencies, and Git changes rather than semantic similarity search.