SGraph MCP Server

An MCP server that gives AI agents instant access to software architecture, dependencies, and impact analysis through pre-computed sgraph models. One tool call replaces dozens of grep/read cycles.

Why?

AI agents discover code structure by reading files one at a time. For a question like "what calls this function?", that means grep, read, grep again, read again... Dozens of round-trips, thousands of tokens, and results that still miss indirect callers.

SGraph pre-computes the full dependency graph. The same question takes one call and returns every caller with type information.

	Traditional (grep/read)	SGraph MCP
"What calls this function?"	Multiple grep + read cycles	sgraph_get_element_dependencies
"What breaks if I change this?"	Manual trace, easy to miss	sgraph_analyze_change_impact
"Show module structure"	ls + read + scroll	sgraph_get_element_structure
Time per query	Seconds (many round-trips)	Milliseconds (cached)
Accuracy	Text matching (noisy)	Semantic graph (precise)

Quick Start

1. Install

git clone https://github.com/softagram/sgraph-mcp-server.git
cd sgraph-mcp-server
uv sync

2. Start the server

# With Claude Code profile (recommended)
uv run python -m src.server --profile claude-code

# With auto-loaded model (skip the load_model step)
uv run python -m src.server --profile claude-code \
  --auto-load /path/to/model.xml.zip \
  --default-scope /Project/src

3. Connect your AI agent

Claude Code / Cursor (.mcp.json)

Create .mcp.json in your project root:

{
  "mcpServers": {
    "sgraph": {
      "command": "uv",
      "args": [
        "run", "--directory", "/path/to/sgraph-mcp-server",
        "python", "-m", "src.server",
        "--profile", "claude-code",
        "--transport", "stdio",
        "--auto-load", "/path/to/model.xml.zip"
      ]
    }
  }
}

Any MCP client (SSE transport)

Start the server with SSE (default):

uv run python -m src.server --profile claude-code --port 8008

Then connect:

{
  "mcpServers": {
    "sgraph": {
      "command": "npx",
      "args": ["-y", "mcp-remote", "http://localhost:8008/sse"]
    }
  }
}

Where do sgraph models come from?

sgraph models (.xml.zip files) are produced by Softagram code analysis or the open-source sgraph CLI tools.

The models represent your codebase as a hierarchical graph:

/Project
  /Project/src
    /Project/src/auth/login.py
      /Project/src/auth/login.py/LoginHandler        (class)
        /Project/src/auth/login.py/LoginHandler/validate  (method)
  /Project/External
    /Project/External/Python/requests                 (third-party)

Each element can have associations (dependencies) to other elements, forming a complete dependency graph.

Tools

The claude-code profile provides 6 tools optimized for AI-assisted development:

Tool	What it does	When to use
sgraph_search_elements	Find symbols by pattern	"Where is the UserService class?"
sgraph_get_element_dependencies	Query incoming/outgoing deps	"What calls this function?"
sgraph_get_element_structure	Explore hierarchy	"What's inside this module?"
sgraph_analyze_change_impact	Multi-level impact analysis	"What breaks if I change this?"
sgraph_audit	Architectural health checks	"Any circular dependencies?"
sgraph_security_audit	Security posture overview	"Any exposed secrets or CVEs?"

The key tool is sgraph_get_element_dependencies with its result_level parameter for controlling abstraction:

result_level=None  ->  /Project/src/auth/login.py/LoginHandler/validate  (raw)
result_level=4     ->  /Project/src/auth/login.py                        (file)
result_level=3     ->  /Project/src/auth                                 (directory)
result_level=2     ->  /Project/src                                      (component)

For the full tool reference with workflows and examples, see SGRAPH_FOR_CLAUDE_CODE.md.

Legacy profile (14 tools)

The legacy profile provides the full original tool set for backwards compatibility:

Basic Operations: sgraph_load_model, sgraph_get_root_element, sgraph_get_element, sgraph_get_element_incoming_associations, sgraph_get_element_outgoing_associations

Search: sgraph_search_elements_by_name, sgraph_get_elements_by_type, sgraph_search_elements_by_attributes

Analysis: sgraph_get_subtree_dependencies, sgraph_get_dependency_chain, sgraph_get_multiple_elements, sgraph_get_model_overview, sgraph_get_high_level_dependencies

uv run python -m src.server --profile legacy

Example Conversation

You: "What would break if I rename the validate() method in auth/login.py?"

Agent calls: sgraph_analyze_change_impact(element_path="/Project/src/auth/login.py/LoginHandler/validate")

Result:
  5 callers in 3 files
  - /Project/src/api/routes.py (2 call sites)
  - /Project/src/middleware/auth.py (2 call sites)
  - /Project/tests/test_auth.py (1 call site)
  Warning: bidirectional dependency with /Project/src/middleware

Architecture

MCP Client Request
       |
[Tools Layer]     src/tools/        -- MCP tool definitions, input validation
       |
[Services Layer]  src/services/     -- Business logic (search, deps, security)
       |
[Core Layer]      src/core/         -- Model management, data conversion
       |
[SGraph Library]                    -- Graph operations (sgraph package)

See ARCHITECTURE.md for the detailed design.

Development

# Run tests
uv run python tests/run_all_tests.py
uv run python tests/run_all_tests.py unit          # Unit only
uv run python tests/run_all_tests.py integration   # Integration only

# Lint
uv run ruff check src/

# Run a single test file
uv run python -m pytest tests/unit/test_collect_deps.py -v

See CONTRIBUTING.md for how to contribute.

About

Built by Softagram using the open-source sgraph library. Licensed under MIT.