NinaCorp

Free · Runs offline · Early test build

A private AI assistant that lives on your computer.

Tiny Agent runs small open-source language models on your own PC. Keep notes, set reminders, ask about what you've saved, and build small websites and apps that are checked in a real browser. Local features need no subscription. Optional conversation uses either paid OpenAI realtime voice, or Spark with local speech using your Spark subscription allowance. Local Whisper recognition requires a one-time setup; Windows dictation remains an explicit alternative. Both conversation options connect your ideas on a reviewable canvas.

Live sketch is a separate all-local way to think visually: tiny CPU speech and a small local model shape a working page, moodboard and HTML preview. Direct controls, undo and reviewed editor handoff are included. This is a rough component-based sketch, not arbitrary functional app generation. It uses no Spark allowance or paid voice API; speech setup is required, and speed and recognition depend on your machine and voice.

Optional Phone Chat brings text conversation to your phone over private Tailscale HTTPS. Pair with an expiring QR and matching-code approval on your PC. Keep TinyAgent open and the PC awake; actions needing approval still wait on the PC. Phone voice and file controls are not included.

Get the test build Read the research

What it does

Notes and reminders

Tell it what to remember. Reminders fire on time, and one that came due while the app was closed is shown as overdue rather than missed.

Answers from your own notes

Questions are answered from what you've saved, and the notes it used are shown beside the answer, so an invented detail is easy to spot.

Websites and small apps

A build counts only when a real draft exists. Pages are opened in an isolated browser and checked against what you asked for; vision-capable models then inspect a bounded rendered preview. Research stays part of the build, and unsupported supplement health claims are blocked unless matching retrieved evidence is cited.

Live sketch

Think out loud and a small local model shapes a rough page as you talk. It keeps a visible note of what you are making, who it is for and ideas parked for later. Thinking aloud changes nothing, and it asks before pulling a parked idea in or changing the goal.

Editor and safe preview

Edit what it drafts and preview it in a sandbox that has no access to your computer.

You stay in control

The model can only use registered tools, and anything risky waits for your approval before it runs.

Your choice of model

Works with Ollama, LM Studio or llama.cpp. Pick a model that suits your hardware: 1.5 to 8 billion parameters run fully on an 8 GB graphics card, and 14 billion needs more video memory.

Honest about small models

Small models are fast, free and private, but they have limits. We tell you where they are, based on what we've measured.

Works well

  • Notes, tasks and reminders
  • Questions about your saved notes
  • Single-page websites, calculators and to-do apps with a 7B coding model

Still hard

  • Long code in one go. Multi-page websites now build page by page with a coding model, but are checked only for completeness and links, not yet for images or styling (RL-016)
  • Complex app logic, such as live filters (RL-012)
  • Code with 1.5B models (RL-010)
  • "Thinking" models often run out of reply budget on long code (RL-015, RL-016)

Use it, or help improve it. Both are welcome.

Just use it

Free, and nothing is shared

Tiny Agent never sends your conversations to us. Sharing is off by default, there's no account, and you never have to contribute to use the app.

Download

Help improve it

Share the moments it got you wrong

Opt in, and you can send a specific conversation where Tiny Agent misunderstood you, after seeing exactly what will be sent. We study the gap between what you meant and what happened, and publish what we learn.

How contributing works

Consumer-grade first

Built and tested on ordinary PCs, not workstation GPUs.

The harness owns the truth

Code decides whatever can be computed. The model is asked only for judgement.

Evidence over claims

33 research-log entries so far, failures included.

Our mission and values →

What you need