TENKA
She drives.
No chat window pretending to help. TENKA is an open-source, local-first desktop AI companion — she reaches into your actual apps and websites, and tells you what she did.
Six senses, alive at once
Voice, vision, memory, browser, apps, and continuous learning — TENKA listens, thinks, remembers, and acts, all on your machine. No cloud. No account. No one watching but her.
Three stacks. The LLM is never asked which one to use.
A deterministic router sends each task to the cheapest stack that can do the job — Playwright for the browser, Terminator for native apps, and vision only when nothing else can see. About 80% of her actions use zero vision calls, which is what every other agent pays for on every action.
A personality that actually evolves
Six traits — warmth, curiosity, directness, playfulness, assertiveness, honesty — each bumped by events and decayed over time by a real engine running locally. The version of her you meet tomorrow won't be the version of her you meet today.
Everything she knows about you, on one SQLite file
A private knowledge graph she builds from your conversations, stored in %APPDATA%/tenka/ as one SQLite file plus a FAISS index. Scrub through time, resolve contradictions, and forget anything instantly. Nothing leaves your machine unless you point her at a cloud LLM yourself.
She plans, she fails, she recovers
TENKA writes a plan, runs it, and when a step fails she replans on the fly instead of giving up — caching the procedure so the next run is free.
About $2 / month for daily use
Most days you stay on free quotas. The only ongoing cost is the LLM provider you choose — Gemini Flash-Lite on the cloud default (~$2/mo), Groq + Cerebras free tiers as a $0 public fork, or Ollama fully offline for $0. There is always a $0 path. Compare that to $20/month for ChatGPT Plus, Claude Pro, or Cursor.
The honest answers
- Is this another cloud assistant?
- No. TENKA is a desktop agent that runs on your machine and acts on your real apps. The cloud is only an LLM provider you point her at — and you can run her fully offline.
- Does it work without internet?
- Yes, with Ollama locally. Quality drops vs Gemini, but every intent — voice, vision, memory, automation — still works.
- How is this different from a screen-recording AI agent?
- Most agents use vision on every action. TENKA's router deterministically picks between browser, native apps, and vision — and prefers the cheap path. About 80% of her actions use zero vision calls.
- Can I add support for my favorite app?
- Yes. Either teach her once (she learns and caches the procedure) or drop a YAML manifest in personalities/apps/. No code changes, no fork required.
- Where does my data live?
- One SQLite file plus a FAISS index in %APPDATA%/tenka/. No telemetry SDK, no analytics. Nothing leaves your machine unless you choose a cloud LLM.
- What if she clicks the wrong thing?
- ESC-hold anywhere aborts immediately. She asks before destructive actions (delete, send, pay, post), and her cursor is visibly TENKA-colored when she's driving.
- What hardware do I actually need?
- Windows 10/11, about 2 GB RAM, any GPU (works on integrated, comfortable on a GTX 1650). Microphone optional for voice.
- What's the difference between TENKA and MateEngine?
- MateEngine is the Unity frontend — the visible desktop character. TENKA is the engine behind her: the brain, memory, and automation. They talk over a local socket, and TENKA can run headless.