Fully local
A custom fine-tuned model runs on your own GPU, so the model never leaves your machine.
Orbital · Guidance · Local
Local AI agents,
engineered for liftoff.
DSKY Software Development builds fast, private AI agents that run on your own GPU. Our flagship, DSKY, turns one plain-English request into finished work across your browser and your desktop.
- ~10×faster than Microsoft UFO + a frontier model on our benchmark task
- ~120tokens per second, served locally
- 16 GBone desktop GPU, no cloud model needed
01 Flagship
Meet DSKY. Your desktop, on autopilot.
Pronounced dis-key and named after the Apollo Guidance Computer’s display & keyboard, DSKY is a desktop AI agent that drives your browser and your computer at the same time. Describe the task in your own words. DSKY picks the right tools, explains its plan, and checks in before risky steps.
Browser + desktop, at once
A custom MCP server lets DSKY operate Chrome and your desktop apps inside the same task.
Reads structure, not screenshots
DSKY works from the page and app element trees instead of guessing from pixels. That’s more accurate and far faster than vision-based agents.
Just say it
Speak the task out loud. faster-whisper transcribes it on-device, and DSKY starts planning.
Plans first, asks before risky steps
You see the plan before anything runs, and DSKY checks in before any step that matters.
Model + framework, tuned together
We optimize the model and the agent framework as one system. That’s how DSKY outruns frontier models on frontier frameworks.
02 Benchmark
One request in. One workbook out.
David’s test task, spoken out loud and run start to finish on a single desktop GPU:
Go to five GitHub repos similar to browser-use and get five pieces of information on each. Then go to Yahoo Finance, choose three companies, and get five financial metrics on each. Then get the five latest videos from the official Carnegie Mellon and UPenn YouTube channels. Use your backend to find the official links first, then compile everything into one Excel spreadsheet.
- GitHub repos
- Yahoo Finance
- CMU YouTube Videos
- UPenn YouTube Videos
DSKY ran the whole job locally at about 120 tokens per second, then opened the finished workbook on its own.
~10×faster on the same task
Wall-clock times measured by David on the task at left. UFO on a frontier model took 10–20 minutes, usually close to 20 (tested with Claude Opus).
03 Projects
The fleet.
Every project taught us something the next one needed. Here’s what we’ve built so far.
DSKY
The desktop AI agent: a Hermes Agent backend and the custom DSKY MCP server. It runs fully local, takes voice requests, and drives your browser and desktop at the same time.
Hermes AgentCustom MCPfaster-whisperLocal LLM
Sherman
An upgraded MCP server running on Claude. Text it a task from your phone and it hands the work to sub-agents that drive multiple windows in parallel.
ClaudeMCPSub-agentsParallel
VEX CAD model
A local GPT-OSS fine-tune that designs VEX robot parts in CAD. When the base model struggled, David ran an SFT + QLoRA pass on his own GPU in about an hour, and the result was a much better CAD model.
GPT-OSSSFTQLoRACAD
DSKY v1
Where it started: an agentic automation framework built on OpenClaw and a custom MCP server. It could open chess.com, play a full game, and email David the results.
OpenClawMCPBrowser automation
Custom Qwen3 fine-tune
David’s first custom local model, fine-tuned on AMD Developer Cloud and sized to run on a single desktop GPU.
Qwen3AMD Developer CloudFine-tuning
autoclaude
An actively maintained AGPL-3.0 remake of AutoClaude, with support for newer Claude models plus Skills and Output Styles.
TypeScriptAGPL-3.0Claude
View on GitHub →
04 Flight log
From “I can build something better” to DSKY.
I told him no, because I believed I could make something better.
-
2025 · The spark
Operator? No thanks.
David’s dad wanted a ChatGPT Operator account, the agent that could drive a Chrome browser. David talked him out of it and set out to build his own.
-
Research
Studying the state of the art
He took apart Microsoft UFO (it drives Windows through UI Automation) and browser-use (it drives Chrome through Playwright and the Chrome DevTools Protocol). The lesson: element-based automation is far more accurate than vision, but element trees are huge and can be slow.
-
Training
Inspirit AI Scholars
He was accepted into Inspirit AI’s AI Scholars program and learned from Stanford and MIT alumni: RAG, framework stacks, neural networks, and how models are trained and used in the real world.
-
Going local
A custom Qwen3
He chose to focus on local models and fine-tuned his first one on AMD Developer Cloud, a Qwen3 model small enough to run on his own GPU.
-
Project 1
DSKY v1
He gave that model a framework: OpenClaw plus a custom MCP server. Its first trick was opening chess.com, playing a game, and emailing him the results.
-
Project 2
Sherman
Next came an upgraded MCP server running on Claude. He could text it from his phone, and it split the work across sub-agents in parallel windows.
-
Project 3
A model that CADs
He fine-tuned GPT-OSS to CAD VEX robot parts, running SFT + QLoRA locally on his GPU in about an hour.
-
Today
DSKY
A Hermes Agent backend and the DSKY MCP server: browser and desktop at once, fully local, and about 10× faster on our benchmark.
05 About
Built by a robotics kid who wanted faster agents.
DSKY Software Development is the company David Blackman started to build the AI agents he wished existed. He’s also robotics president and lead CAD designer for Fenwick High School’s VEX program, so he approaches agents like robots: measure them, find the bottleneck, rebuild, repeat.
That’s why DSKY tunes the model and the framework together, runs locally instead of renting someone else’s cloud, and reads app structure instead of squinting at screenshots. Each project in the flight log solved a problem the one before it exposed.
Toolbox
Local LLMsFine-tuningSFTQLoRA
Qwen3GPT-OSSMCP serversHermes Agent
OpenClawfaster-whisperAMD Developer CloudClaude
CADVEX Robotics
06 Early access
Be first in line for DSKY.
DSKY is in active development. Join the list and we’ll email you when early access opens.
Collaborations, press, or just want to talk agents? [email protected]