AI Partner
The partner
that knows you.
Works with any AI you already pay for, or fully local models. It all runs on your own machine, so your data never leaves.
What it feels like
You teach it once.
It still knows tomorrow.
Switch engines mid-conversation. The partner keeps the thread.
The model underneath can change. The mind that knows you does not.
The shift
The partner you needed
didn't exist yet.
Three sentences for three eras. The middle one is now.
You rented amnesia. Every session started from zero, every model forgot you the moment the tab closed, every morning began with the same re-explaining.
Same partner. Every model. The memory survives every swap. A mind that knows you, sitting on top of whichever model is best this quarter.
Memory and language stop being two layers and fuse into one. The partner becomes yours, outliving every provider you use.
What it does
Three things it does
that nothing else does together.
Check it yourself
open coreFour things you can verify in the repo.
No self-reported dashboards. Every claim here points at code you can read and a suite you can run.
Open protocol·open SDK·closed engine·your keys, your machine
The five components
Five named components.
Each with one guarantee.
What each does, what it guarantees, what you see. How it works inside is the part we keep to ourselves.
The engine is yours
Works with what
you already pay for.
Zero inference of our own. No middleman between you and the model. The partner runs on your subscription, your keys, or your own hardware.
The subscription you already pay for
Point it at your Claude, ChatGPT, or Kimi plan and it uses what you already have. No second bill, no per-token surprise.
Fully local, auto-picked for your Mac
Run entirely offline. The partner picks the best local model your RAM can hold, so you get the strongest option your machine allows without tuning a single flag.
Bring your own keys
DeepSeek, Kimi, Grok, Gemini, Qwen, OpenRouter, or any OpenAI-compatible endpoint. Your key, your quota, your choice of provider.
Per-pane control that persists
Set the engine per pane with /model. The choice sticks across restarts, so each workspace keeps the brain you gave it.
What each one needs from you, a subscription, a key, or nothing at all: the full list →
For Claude Code users
Fixes what you hate
about Claude Code.
It forgets the moment the window compacts.
A persistent substrate holds the thread. Compaction stops meaning amnesia, because the memory does not live in the context window.
Every session starts from a blank slate.
The partner picks up where you left off. What it learned about your repo and your decisions is still there tomorrow morning.
Swap the model and you lose everything.
The model is a component you can change. The mind that knows you stays put through every swap.
You re-paste the same context every day.
Teach it once. It keeps your project, your patterns, and your taste without being re-fed them each session.
Open source
The core is open.
The app is optional.
The substrate, the protocol, the CLI, and the test suite are AGPL and free forever. It is plain Node and runs anywhere Node runs: Linux, a server, a headless box. The Mac app is a one-time purchase that puts a native face on the same core.
Then run troth init. No account, no server, your data stays on your machine.
Who it's for.
Engineer
You want one partner that knows your codebase and stays through model swaps.
Founder
You want an AI that compounds your team's context, not one that resets every Monday.
Researcher
You want a partner that remembers every paper, every experiment, every contradiction.
Shipping to a team? For your company →