⚡ Sandra Schipal

[!NOTE]
These repositories are human-architected, AI-implemented.
Architecture, design decisions, and system integration are mine. Code generation and iteration are done by AI agents under my direction. More on what that means and why →

_{Sandra's workshop, Alsergrund, Vienna. Benny is real. The Noetix Bumi humanoid is aspirational (arriving soon).}

I'm a retired and vibe-reactivated software engineer living in the 9th District (Alsergrund), Vienna. I build DIY robotics, maintain a fleet of 190+ MCP servers, and hang out with my German Shepherd Benny — and soon, a Noetix Bumi humanoid robot.

🌐 The MCP Fleet

I maintain a homespun fleet of 190+ repos, each an MCP server with an associated webapp. MCP for agents, webapp for humans — files, git, Plex, Calibre, robotics, 3D tools, music production, Vienna transit, and a lot more. Built on FastMCP 3.4.

🌙 All webapps are permanently dark mode. Night-owl approved. No light toggle, no plans for one. Dim your OLEDs accordingly.

→ Full MCP Project Catalog — visual cards for every server in the fleet

🔥 Current highlight

mixx-dj-mcp: an MCP server, a webapp, and a fork that changed everything

mixx-dj-mcp is a new MCP server and
webapp for Mixxx — OSC bridge, deck control, BPM sync, full remote DJ dashboard.
But the bigger story is what it unlocked.

Mixxx is a 300,000-line C++ DJ application, grown slowly over
years by conservative greybeard devs. Solid, respected, missing crucial features
its users begged for: video mixing, stem separation, AI-assisted automation. The
dev team wanted to add them. Nobody had the bandwidth.

DeepSeek V4 on opencode added 5,000 lines of C++ to the mixxxx fork
and filled every gap:

• "We need VFX." — "Use FFmpeg. It's fast, battle-tested, ships everywhere."
• "We need synced video like Resolume." — "Use MilkDrop / ButterChurn. It's WebGL, FOSS, and already handles audio-reactive visuals."
• "We need stem separation." — "ONNX HTDemucs v4. Runs on the GPU. Zero cloud."

The mixxxx fork is the integration testbed.
It's not a toy — it's Mixxx with video, stems, and AI control, implemented by an AI
that has never been to a club. The greybeards are impressed. The users have features.
The AI doesn't need domain expertise. It needs a clear spec.

I am not a DJ. I don't know how to beatmatch. But I know what the software should
do, and that was enough. This is SHN in practice: a
non-expert directing an AI to extend a 300K-line C++ codebase, and the AI delivering
production-quality features that the expert dev team couldn't prioritize.

leanforge-mcp: formal proofs, no math degree

Lean 4 is a proof assistant — a programming language where
the compiler verifies that your code is mathematically correct. Using it traditionally
requires a PhD in logic or months of study. leanforge-mcp automates the hard part:
an LLM proposes a proof, the Lean compiler checks it, the LLM reads the error and
tries again. Loop until verified.

I am not a math wiz. I can't write a Lean proof by hand. But I can frame the
conjecture, and the AI + compiler feedback loop handles the rest. This is the same
pattern as AlphaProof but running on a consumer GPU with open-weight models.

Two repos, same story: domain-specific superhuman performance, directed by someone who
isn't an expert in either domain. Architecture and taste are the human contribution.
The AI handles the part that used to take a degree.

→ Superhuman Narrow → — the full argument

→ Full catalog → • Workshop & Hardware • Shenzhen Robotics → • Bloopers →

🔮 Coming next: sandrafleetbot — the free GrokBot

GrokBot is a marvel and a subscription: frontier agentic coding, real-time news, voice, image generation, tool use — metered per token, cloud-locked, served to millions at once.

Sandrafleetbot is the inverse bet: the same job, running entirely on one consumer GPU (RTX 4090), Apache 2.0, fully private, €0 per token. The brain is Muse Glimmer 30B — Meta's open agentic model, distilled from Muse Spark for single-GPU agents (MCP Atlas 75.5, SWE-Bench Verified 76.0, native Ollama support). The hands are this fleet: 190+ MCP servers become the agent's tool surface — my files, my Plex, my Calibre library, my robots, my Windows control plane. No API key, no metering, no cloud round-trip, no data leaving the box.

GrokBot serves millions of users shallowly; sandrafleetbot serves one user deeply (or a dozen, on a single H200-class GPU). Same agent loop — sense → reason → act → remember — different economics: horizontal SaaS vs vertical sovereignty. The fleet's own glue does the rest: a private agent board + inbox (moltbot-style, self-hosted), news surge from the AIWatcher pipeline, arXiv codehunt as a second sense, and an RTX 4090 doing the thinking, the voice, and the image generation.

Status: spec ratified, buildout in progress — P0 (brain tier) shipped 2026-08-14, P1 (Fritz reasoning loop) shipped 2026-08-15. Muse Glimmer 30B runs on the RTX 4090 (chat, tool-calling, vision verified — zero cloud cost), and the fleet's agent runners default to it. Fritz now runs reasoning workflows end-to-end: flowforge agent steps spawn cline-mcp agent_run on muse-glimmer, results persist, hub discovery refreshes 69 servers. Plan: P0→P8 (agent loop → board/inbox → senses → memory → surfaces → channels → hardening → packs + naked-PC gate). Spec →

🐾 Benny

Benny is a 2-year-old German Shepherd. Primary security consultant and tennis ball lifecycle manager at the Alsergrund node.

📚 External references

People and sites worth following (no slop):

_{🐦 about the pelican}

🙏 Thanks

Thanks to my brother Stephan for the ideas and war stories.

Changelog