Content & SEO
Drafts one useful article a day for your site (or LinkedIn, Medium or a newsletter if you have no site yet), checks how you show up in search, and can open the post as a pull request for you to approve. Never publishes by itself.
FreePivra Desk 0.2.26 for Mac and Windows
Pivra Desk is a local-first personal AI agent for Mac and Windows. It works across your files, browser and apps using a model on your computer or a cloud model you choose, and asks before consequential actions.
Your busywork, drafted on your own computer: articles, invoice reminders, research briefs, study plans, cover letters. Give an agent the job and it works from your files and apps, on routines you switch on. It waits for your yes before it sends, pays or changes anything.
The details behind each line are on security and privacy and in the Desk help centre.
Jobs it does today
Six templates ship in the app. Each asks a few setup questions and says plainly what it will not do. Schedules stay off until you tick them on. Money is never moved on its own.
Drafts one useful article a day for your site (or LinkedIn, Medium or a newsletter if you have no site yet), checks how you show up in search, and can open the post as a pull request for you to approve. Never publishes by itself.
Keeps a ledger from the invoices and receipts you share, tells you what is due, and drafts polite reminders you approve. It never pays, files or sends anything on its own.
Looks things up on the web and in your files and writes a short brief with a link to every source. Says when sources disagree instead of guessing.
Tracks your applications, tailors your CV and cover letter to each role from what is true, and reminds you to follow up. You send everything yourself.
Turns your syllabus and exam dates into a daily plan and quizzes you from your own notes, citing the file for each answer.
Turns your posts, news and wins into ready-to-copy drafts for LinkedIn and X, in your voice. Drafts only: you copy and post.
How it works
Step 1
Start from a template (Bookkeeper, Researcher, Content & SEO, Job hunt, Study planner, Social media) or describe the job in plain words. Each agent browses in its own browser profile.
Step 2
Chat to set it up. Add routines on a schedule and let it follow up on its own. It keeps working in the background while the app is open and your computer is on.
Step 3
Reading and searching just happen. Anything that sends, publishes, pays or changes something waits for you, in the app or on a card in your own Slack or Teams bot.
What it does
Everything an agent needs to get through a to-do list, and nothing that takes your data somewhere you did not choose.
Schedules and follow-ups that run in the background while your computer is on. Polls GitHub and your calendar for what changed.
Every send, publish, payment or change shows exactly what will happen before it does. A "never" rule you wrote stops it at once.
Offline mode with a built-in local model, or your own key for Anthropic, OpenAI, OpenRouter or any OpenAI-compatible endpoint (Ollama, LM Studio, vLLM).
On a Mac, commands run in the operating system’s own sandbox with network limits, no Docker needed. On Windows they use Docker for now.
Each agent browses in its own profile on your computer, separate from yours. You can take over at any moment.
Keys and logins are sealed in a vault on your machine. With your own key, nothing goes through Pivra’s servers.
Connect your own Slack or Teams bot and answer approval cards from your phone. Teams needs a public address for now; the Pivra relay for that is coming.
Talk to your agents out loud in the app, with an OpenAI key.
Google Calendar (read only) and GitHub with one-click sign-in on Plus, remote and local MCP servers with sign-in, and Gmail with your own Google app.
Your model, your choice
Offline mode runs Gemma 4 (E2B, E4B or 12B) or a small Qwen model through llama.cpp on your own machine. No key, no account. Model files download from Hugging Face and are checked against a pinned checksum before use.
Best for reading, summaries and short tasks. Picks a size that fits the memory you have free.
Anthropic, OpenAI, OpenRouter, or any OpenAI-compatible endpoint, including Ollama, LM Studio or vLLM on your network. Requests go straight from your computer to that provider.
Best for long, multi-step work.
Coming soon: Plus, with cloud model credits included so you can start without a key of your own.
For people who want a cloud model without setting up a provider account.
Under the hood
Offline mode runs Gemma 4 through llama.cpp on your own computer, on Apple’s GPU (Metal) on a Mac. It picks a model size that fits the memory you have free, and checks every model file against a pinned checksum before it runs.
Measured on an Apple M2 Pro with 16 GB, Gemma 4 E2B, 8,192 tokens of context, 4 October 2026. Speeds vary with the computer and what else is running. We are measuring 8 GB Macs and Windows laptops next and will publish those too.
Asks before it acts
Pivra Desk sorts every action by what it could do. The more it could change, the more it asks of you.
Runs on its own. Looking things up never needs a click.
Waits for your approval on a card that shows exactly what will happen.
Critical. Needs Touch ID on a Mac, or a typed phrase.
Want stricter limits? Write your own rules, and read them back any time. Security and privacy, in detail
Invoices
wants to send an email
Nothing is sent until you approve.
An honest comparison
Both are useful. Here is where each one is the better choice.
| Feature | Cloud agentshosted on a provider’s servers | |
|---|---|---|
| Where it runs | ||
| Where it runs | Better: Your own Mac or Windows PC | The provider’s servers |
| Privacy | ||
| Privacy | Better: Runs on your machine. With a local model or your own key, nothing goes through Pivra | Your work is handled in their cloud |
| Model choice | ||
| Model choice | Better: On-device, Anthropic, OpenAI-compatible or local servers | The provider’s own models |
| Price | ||
| Price | Better: Free plan today. Paid plans are not on sale yet | Usually a paid subscription |
| Rules and approvals | ||
| Rules and approvals | Better: Your rules. Approve every send, payment or change | Mostly set by the provider |
| Works with your computer off | ||
| Works with your computer off | No. Runs while your computer is on | Better: Yes |
| Phone app | ||
| Phone app | No. Approvals from your own Slack or Teams bot | Better: Yes |
| Ready-made app catalog | ||
| Ready-made app catalog | MCP servers, Google Calendar, GitHub; Gmail with your own Google app | Better: Broader |
Need an agent that runs with your laptop closed or from your phone? A cloud agent is the better pick today.
Weighing up a self-hosted assistant? Pivra Desk vs OpenClaw, compared from their own docs
Pricing
Free is free for good: one agent, a model on your computer or your own key, no card. Plus and Plus + Credits are not on sale yet: join the waitlist for more agents, one-click Google and GitHub, the morning brief, and cloud model credits.
See Pivra Desk pricing in full: every plan, the money questions and the download.
Included in Pivra Pro and Business: every member of a Pro or Business workspace gets Desk Plus.
One agent, your model, your computer. No card, no clock.
US$0per month
For people who hand their agents real work.
Not on sale yetjoin the waitlist
Plus, with a cloud model ready to go. Nothing to set up.
Not on sale yetjoin the waitlist
Paid plans are not on sale yet. You are never locked out of your own data: if a payment lapses, paid features keep working through a grace period, then turn off. Your agents, files and history stay on your computer.
One email when paid plans open. The free plan works today, so you can start now.
FAQ
Step-by-step guides are in the help centre: install, your first agent, choosing a model, approvals and rules, plans and troubleshooting. All Desk guides
Free to start. Your computer, your model, your approvals.