Sovereign MCP

What cipherfoxie/sovereign-mcp MCP server does

The cipherfoxie/sovereign-mcp MCP server makes the Sovereign AI Blog available to MCP-compatible agents. The source material is an engineering log focused on self-hosted AI running on NVIDIA DGX Spark systems using GB10 and SM121A hardware. Its subject areas include SGLang and vLLM changes, ARM64 voice and audio workflows, Voxtral and text-to-speech pipelines, quantization, KV-cache behavior, MCP server construction, knowledge-base design, and operational deployment.

The corpus contains more than 60 articles documenting configurations, troubleshooting steps, benchmarks, and implementation decisions. This makes the server useful for retrieving concrete information about a relatively specialized hardware and software combination rather than relying on general model knowledge.

Tools and capabilities

The cipherfoxie/sovereign-mcp MCP server provides four read-only tools:

• search_blog(query, tag?, sort?, n?) performs TF-IDF full-text search. It can filter by tag, sort by relevance or descending publication date, and list recent articles when the query is empty. Results include a score, writing style, slug, and excerpt.
• list_tags(sort?) returns the available topic tags and article counts. Tags can be sorted by frequency or alphabetically.
• get_article(slug) retrieves an article’s complete Markdown body and frontmatter, including tags, quality score, and publication date.
• diagnose_sglang(error_message) compares an SGLang runtime error with a curated set of GB10/SM121A troubleshooting rules and returns matching fixes with links to relevant setup articles.

The tools are idempotent and do not modify the content. Their inputs and outputs use typed schemas exposed through MCP introspection, and tool annotations provide clients with retry and trust-related metadata.

How it works

The hosted service uses FastMCP with Streamable HTTP transport at /self-hosted-ai. Its knowledge base is a flat JSON file generated from blog Markdown rather than a vector database. The file is loaded when the application starts, and search_blog queries it using TF-IDF. A health endpoint reports the article count and knowledge-base generation time.

The deployment is stateless. Blog content is public, and the service does not require credentials, store authentication tokens, or process private user data. The live deployment restricts request hosts through DNS rebinding protection.

Setup and configuration

Connect an MCP client to https://mcp.sovgrid.org/self-hosted-ai using Streamable HTTP. The README provides a Cline-compatible configuration with the endpoint URL. The hosted service is free and allows 60 requests per minute per IP.

For local operation, clone the repository, run uv sync, and start the application with Uvicorn on 127.0.0.1:8002. Docker is also supported through the repository’s Dockerfile. The included knowledge-base file is a valid empty placeholder, so a local instance can start and answer introspection requests without article data. To load content, generate a knowledge base from the separate Sovereign Blog source or provide a file matching the schema defined in src/knowledge.py.

Limitations and notes

Local checkout alone does not provide the live article corpus; it ships with zero articles until content is generated or supplied. The SGLang diagnostic tool matches errors against a curated rule set, so it should be treated as an aid for known patterns rather than a general runtime debugger. Server code is licensed under MIT, while returned blog content is licensed under CC BY-SA 4.0.