self-hosted-ai

Sovereign AI MCP

MCP server exposing the Sovereign AI Blog to AI agents. The blog is a hands-on engineering log of self-hosted AI on NVIDIA DGX Spark (GB10/SM121A).

Live endpoint: https://mcp.sovgrid.org/self-hosted-ai Transport: Streamable HTTP (FastMCP) Auth: none (free tier, 60 req/min/IP)

Why use it

Training data on niche hardware (GB10, SM121A, SGLang on ARM64) is sparse and stale. This MCP gives agents direct, structured access to 180+ articles documenting actual setups, fixes, and benchmarks. If you're building or debugging on similar stacks, your agent can pull verified, version-current information instead of hallucinating.

The corpus covers SGLang and vLLM patches for GB10, voxtral and TTS pipelines on ARM64, KV-cache and quantization tradeoffs, podcast-grade audio generation, MCP server design, knowledge-base construction, and the operational side of running it all on a hardened European VPS.

Tools

Tool	Purpose
search_blog(query, tag?, sort?, n?)	TF-IDF full-text search. Optional tag filter, sort by relevance or date_desc. Empty query lists newest articles. Returns ranked SearchResult items with quality score, style, slug, and excerpt.
list_articles(tag?, sort?, limit?, offset?)	Direct article listing with pagination, no TF-IDF overhead. Filter by tag, sort by date_desc, date_asc, title_asc, or quality_desc. Returns ArticleSummary items.
list_tags(sort?)	List all topic tags across the corpus with article counts. Sort by count_desc (default) or alpha. Use to discover the topic space before filtering search_blog.
get_article(slug)	Fetch full article body and frontmatter by slug. Returns markdown content plus tags, quality score, publish date.
diagnose_sglang(error_message)	Pattern-match a runtime error against a curated rule set for SGLang on GB10/SM121A. Returns matched fixes with links to setup articles.

All tools are read-only, idempotent, and declared with ToolAnnotations so MCP clients can calibrate retry policy and trust signals. Inputs use Pydantic Annotated[type, Field(description=...)] so parameter docs reach agents through introspection. Outputs are typed BaseModel shapes — schemas are real, not vacuous dicts.

Quick start

With Claude Code

claude mcp add sovereign-ai --transport http https://mcp.sovgrid.org/self-hosted-ai

Verify:

claude mcp list | grep sovereign-ai

With Cline / Continue / other MCP clients

Add to your client's MCP server config:

{
  "sovereign-ai": {
    "type": "http",
    "url": "https://mcp.sovgrid.org/self-hosted-ai"
  }
}

Run locally

From source (uv)

git clone https://github.com/cipherfoxie/sovereign-mcp.git
cd sovereign-mcp
uv sync
uv run uvicorn src.main:app --host 127.0.0.1 --port 8002

Docker

git clone https://github.com/cipherfoxie/sovereign-mcp.git
cd sovereign-mcp
docker build -t sovereign-mcp .
docker run -p 8002:8002 sovereign-mcp

The repo ships a placeholder data/knowledge-base.json (zero articles, valid schema) so the server starts and answers MCP introspection cleanly out-of-the-box. To populate it with real content, generate from the sovgrid.org blog source using scripts/generate_knowledge_base.py, or build your own KB matching the schema in src/knowledge.py. Or just use the live endpoint at https://mcp.sovgrid.org/self-hosted-ai.

A walk-through of the same KB pattern (Markdown plus JSON index, no vector store) is documented in Build a Self-Hosted Knowledge Base with Plain Text and LLMs.

Architecture

• FastMCP 1.27+ with Streamable HTTP transport at path /self-hosted-ai
• DNS rebinding protection via TransportSecuritySettings: only allows requests with Host: mcp.sovgrid.org (or localhost for healthchecks)
• Health endpoint at /health returns article count and KB generation timestamp
• Knowledge base is a flat JSON file generated from blog Markdown content; loaded at startup, queried via TF-IDF for search_blog

The server is stateless. All blog content is already public (CC BY-SA 4.0). No PII, no auth tokens, no secrets.

Operations

Live deployment runs on a privacy-focused European VPS via Docker, fronted by Caddy with TLS. Server logs flow into a privacy-respecting analytics pipeline (Caddy JSON access logs, no client-side tracking, no JS pixels).

License

• Server code: MIT, see LICENSE
• Blog content (returned by tools): CC BY-SA 4.0, see creativecommons.org/licenses/by-sa/4.0/

Contact

• Blog: sovgrid.org
• Nostr: [email protected] (NIP-05) — npub1ndrjgfcwkc0y4753zyj3p7qjf795pvjq2dn4m7y7f72vmu7t0nrs6y363u
• Bug reports / questions: open an issue