Architecture-aware AI coding, on your own machine

What it is

Softagram Analyzer MCP packages the Softagram code analysis engine and an MCP server into one container image. You run it on your workstation or build server. Your AI assistant asks it to analyze a repository; the container builds a dependency model of the code and then answers the assistant's questions about that model.

Instead of reading file after file and searching text to guess how the code fits together, the assistant asks the model directly — which keeps its answers grounded in the real structure of your code.

Why run it locally

• Your source code stays on your machine. The container reads the repositories you mount into it, read-only, and keeps the results inside the container or in a volume you choose. It does not upload source code or analysis results anywhere.
• No phone-home. The image has no telemetry and no licence check; it never contacts Softagram.
• Works offline. The analysis runs with networking switched off entirely. Only the package vulnerability checks need to reach a package registry: the npm audit for JavaScript projects (registry.npmjs.org) and NuGet for .NET projects. They send the locked package names and versions, not source code; without network access the rest of the analysis still runs, only those findings are missing.
• Works inside your own security boundary. The MCP endpoint is meant to be reached from your own machine only; the setup instructions bind it to localhost.

Questions your AI assistant can answer

• What depends on this class, function or file? (incoming dependencies)
• What does this module use? (outgoing dependencies)
• What could break if I change this? (direct and indirect dependents, grouped by file)
• What is inside this directory, file or class — without reading the source?
• Where is something defined? (search by name or regular expression)

Analysis results can also be exported as a model file for use with the open-source sgraph MCP server and the sgraph library.

Languages

The release build is verified on a multi-language sample project covering C and C++; C# and .NET project files; Java and JavaFX FXML; Python; JavaScript, JSX (React) and TypeScript; HTML; Go; Scala; PHP; and QML.

The image also contains analyzers for further languages, such as Kotlin, Rust, Ruby and Swift, and reads package manifests and lockfiles (npm, Yarn, pip, Maven, Gradle, NuGet). These are not part of the release test yet, and the depth of the analysis varies by language — ask us about your stack.

Works with your MCP client

The server speaks MCP over Streamable HTTP (the current standard for MCP servers reached over HTTP) and can be switched to the older SSE transport. Setup instructions are provided for Claude Code; other MCP clients that can connect to an HTTP or SSE server can use the same endpoint.

What you need

• Docker: Docker Desktop on macOS or Windows, or Docker Engine on Linux.
• An x86-64 (amd64) or ARM64 machine.
• About 3 GB of disk space for the image, plus space for the analysis results.
• An MCP-capable AI assistant, for example Claude Code.

Getting access

Softagram Analyzer MCP is available to Softagram customers and partners. Partners get access to the container registry, setup instructions and release notes in the Softagram partner portal.

Pricing and licence terms are agreed per customer. Contact us for an offer.