The questions people ask before they install, answered short. Each one links to the page that holds the detail.
Does it run locally?#
Yes. Taskuary is an app on your own machine, not a service you sign up for. It keeps everything
in a SQLite file under ~/.taskuary/ and serves its interface at http://127.0.0.1:7787, for you
only. There is no Taskuary account and no Taskuary server: nothing you connect passes through us.
Setup on Windows is one file, Taskuary.exe.
Anywhere with Python 3.10+, it is pip install taskuary. Docker works too. From there, Set up
Taskuary walks four steps: who you are, one AI, one work source, and a first read. See
Start here.
To look before connecting anything, taskuary --demo runs the real interface over invented work
and cannot reach a real system.
Does it need internet access? What can the browser reach?#
Taskuary itself only talks to the services you connect — your mailbox, your chat, your trackers and the AI you chose. With a local model, the AI calls never leave the machine.
When an agent needs a website, it drives a real browser on your machine (the optional
agent-browser tool), so it reaches whatever your machine can reach, over your network and VPN.
The page shows live beside the agent's terminal. It never types a password or a 2FA code: you
press Take over, sign in yourself, and hand it back. Sites it should only read can go through
a search or reader connector instead, with no browser at all. See
When the agent opens a browser.
Do I need to buy tokens?#
No Taskuary tokens or seats exist: the app is free and open source (MIT). The AI is the only thing that costs, and you bring your own, in whichever of these you already pay for:
| You have | Taskuary uses |
|---|---|
| An API key (Anthropic, OpenAI, Azure OpenAI, OpenRouter) | the key, billed per use by that provider |
| A ChatGPT plan | Sign in with ChatGPT, billed to the plan under a weekly cap you set |
| A coding CLI subscription (Claude Code, Codex and others) | the CLI, with a light model for routine mail |
| None of these | Ollama or another local model server, at no cost per message |
Triage reads many messages and wants a small cheap model; coding wants a strong one. They can be different. See Choosing the AI setup.
Will you sign a BAA? Can it handle patient data?#
There is no Taskuary service holding your data, so there is nothing for us to sign a BAA for — the data stays on your machine and in the systems you already use. The agreements that matter are the ones you hold with whoever does receive data:
- The AI provider. A hosted model receives the text of what it reads. If that can include PHI, use a provider and account under your organisation's BAA (Azure OpenAI and the enterprise tiers of the major providers offer one), or a local model, which sends nothing out.
- Your mail and other systems. Taskuary reads them with your own sign-in, so the BAA you already hold with them still applies.
Before a prompt reaches a hosted model, Taskuary replaces recognisable credentials with placeholders. It does not remove PHI: whether a given setup meets your HIPAA obligations is a decision for your compliance team. See What leaves your machine.
Can it be trained on our software?#
Not by retraining a model — it learns the way a new hire does, by being given the material and corrected on the job:
- Your documents. The Knowledge base indexes folders on the machine and SharePoint libraries (manuals, SOPs, policies). Agents, the Assistant and reply drafts are handed the passages that bear on the work. See Connections.
- Your systems. Connect what has an API; for what only has a web screen, the agent works it in the browser while you watch.
- Your jobs. The first time an agent does a kind of job, Taskuary drafts a playbook — the steps, what it may do alone, what to ask first, what counts as done. Approve it and the next job like it follows it. See Profiles and playbooks.
- Your judgment. Each verdict you correct becomes a lesson for the next message like it. See How it works.
All of it lives in files on your machine that you can read, edit or delete. Taskuary trains no model on it.