M.A.L.T.Y
An AI harness, and a helper with a face that lives on your desktop.
The repository is private while the project is in motion.
- Roles
- Coding · research · light · companion
- API
- OpenAI-compatible
- Memory
- Vault with hybrid search
- Budget
- 12 GB GPU + 6 GB for the rest
M.A.L.T.Y stands for Multi-model Assistant, Learning, Technology, Yield. It is two things in one project: a harness that decides which model answers what, and a small character who lives on your desktop and mirrors what the harness is doing.
Everything runs on one machine. The GPU holds the model, a fixed memory budget covers everything else, and online models are optional extras you switch on with an API key or a ChatGPT sign-in.
Clients
VS Code
Continue + status bar
Web UI
orb, chat, memory graph
Desktop companion
Electron sprite
Gateway
OpenAI-compatible API
model@effort per request
Model router
four roles
Ollama or online model
local first
Memory
Vault of notes
Obsidian style
Hybrid search
plus link-graph expansion
Research agent
cited, saved back
What it does
The parts worth talking about
One endpoint, many models
An OpenAI-compatible gateway with four roles and a model router. A client picks the model and effort per request (for example `helper-coding@high`), and the router can fall back to the local model when an online one is unavailable.
Memory that grows
Long-term memory is a vault of linked notes with hybrid search, link-graph expansion, write-back of what was learned and a lint pass to keep it tidy.
A research agent
Searches the web, reads pages through a safe extractor, answers with citations and can save the result straight into the vault.
Built to stay inside the machine
A Windows job object caps memory, a resource monitor watches the GPU and Ollama, and a `doctor` command says what is missing on a new PC. Setup is scripted end to end.
In the editor
Continue handles chat, edits and autocomplete against the harness, and a small VS Code extension I wrote puts a live status in the status bar: writing code, thinking, reading, searching, listening.
A companion with a state machine
The desktop sprite reacts to the harness: thinking, coding, reading. Left alone she wanders, sits, naps and climbs walls; she dances in headphones when music plays. Click her or press Ctrl+Alt+Space to talk.
Stack
Built with
- Python
- FastAPI
- Ollama
- fastembed
- WebSockets
- Electron
- JavaScript
- VS Code extension
- Continue
- pytest