Domain-specific tool servers for people running their own AI infrastructure. Free, no signup, no KYC, no telemetry on tool calls. One server live today, more on the roadmap.
/self-hosted-ai
Engineering log + SGLang diagnostics for people running LLMs on their own hardware.
Four tools: search_blog, list_tags, get_article, diagnose_sglang.
/lightning
V4V tools, Lightning zaps, L402 paywall helpers. Lets agents send and verify sat-denominated micropayments for premium tool calls.
/voxtral
TTS prompt validation, expressivity-rule lookups, podcast-script linting.
Built on the production-tested rules in kb/forbidden-markup-voxtral.
Each MCP gets its own path under mcp.sovgrid.org/<topic>. Same auth model,
same rate limits, same trust posture. New ones can be added without breaking existing clients.
MCP (Model Context Protocol) is the open standard from Anthropic that lets AI agents call external tools instead of just generating text. When Claude Code, OpenClaw, Cursor, or any MCP-compatible client needs current information or a domain-specific operation, it asks an MCP server for it.
The servers here focus on one thing: helping people who run their own AI infrastructure not waste hours on problems that have already been solved and documented. Read the MCP specification
You bought a DGX Spark, a used 3090, a Strix Halo, or you're scoping a build. You're trying to run an LLM at home and have hit your third weird config error this week. Generic tutorials don't cover GB10 + SGLang + EAGLE. You want answers from someone who actually did the work.
You're building with Claude Code, OpenClaw, Cursor, or your own MCP-aware agent. You want your agent to know what works on real DGX Spark hardware before it suggests a config that OOMs on the first batch. A tool call beats fifty Stack Overflow tabs.
You won't paste configs into a cloud chatbot for fear of data leaks. You want infrastructure knowledge from someone who shares your values: privacy-first, no telemetry, no platform that can deplatform you. You pay in sats when you pay at all.
Free tier today. One server live, two scoped. No rate-limit beyond fair use (60 req/min/IP), no signup.
Premium tools land in Phase 3 behind an L402 / Lightning paywall. Planned: the Lightning MCP itself, deeper diagnostics across more hardware classes, full-text RAG over private engineering notes, and structured config-generation for sovereign-AI stacks. All sat-denominated, no subscriptions, no accounts.
Read the hybrid strategy article for context on why this is structured as agent-first, not human-first.