Your AI workspace, on your terms.
Chat, write code, and work with files in one place. Choose the model and decide what its tools may do.
Windows beta · Linux CLI setup
Getting started
These tools run in your browser. Your entries are not sent anywhere.
Before you download
Use this to plan, not as a benchmark. Setup measures the actual model fit. Allow about 8 GB free disk space and an internet connection for installation.
01 / Your workspace
The workspace keeps your conversations, models and agents in one place.
Inside the workspaceIt stays on this machine.
Your chat history is stored locally, in a folder you can open. Cloud models and connected tools send the content they need to the services you choose.
Features
01 · Conversation
Pin the model you want. Your history stays on your disk, and tools can do only what you allow.
The Instructions box ships empty. Your instructions are sent as written. Tools add the descriptions a model needs to use them. You can inspect the request in the app under “What the model actually receives”.
You choose a permission mode for each conversation, from Plan (nothing changes) to Bypass (everything runs). Read the handbook.
02 · Assistants
Each agent works from instructions in a file you own. Choose its model and tools to fit the work, then set its permissions.
Instructions are plain-text files you can edit. Each agent has a model, a colour and a configured tool set. Review those settings before allowing an agent to act on your files. Read about the controls.
03 · Spending
Paid calls stay off until you set a monthly budget. Local inference has no provider fee.
Usage is recorded from the provider's own figures. PAVONA warns you at four fifths of the limit and stops paid calls when you reach it. Provider billing can lag, so check your provider account too.
04 · Discovery
Discover models from Ollama and OpenRouter without waiting for a new app release.
Local model names come from Ollama, and cloud names come from the live OpenRouter catalogue. Availability depends on the provider and your account. The included prompt-injection probes are a small mechanical check, not a certification.
PAVONA for companies
Start with a workflow your team already runs. Agree who owns it, what data it can use and how you will judge the result.
Pilot scope, pricing and support are agreed before work begins.
01 / DevAd
DevAd builds in a separate copy of the code. You decide what success means before a change is promoted.
Evaluate front ends, back ends and client-specific software against your acceptance criteria.
02 / VERIDITAS
Use it for defensive checks and authorised testing. Written scope and asset-owner permission come first.
Agree the assets, permitted actions and evidence to retain before testing begins.
03 / Support
Set up local assistants for your team. You review each problem report before anything is sent.
The app composes reports with version and system details for you to inspect.
PAVONA is a Windows beta. Its enterprise policy catalogue, provider restrictions and audit records cover specific execution paths; they are not a complete managed enterprise service. Central administration, SSO, tenant isolation, contractual support and deployment requirements need a separate review for your use case.
Security testing requires the asset owner's authorisation. No certification, compliance status or third-party partnership is claimed here. A pilot should produce evidence and a list of remaining gaps, not an assumed approval.
Privacy
Local models run on your hardware. Cloud models and connected tools send data to the services you configure.
Web research, model downloads, cloud providers and other connected tools use the network. Keeping files on your own disk does not protect them from other people or software on the same machine.
The update check reads the release list every six hours while the app is open. Updates install only with your confirmation. What leaves your computer.
From download to first conversation
One file, 48 MB. Get the current installer and its checksum from the release page.
Double-click it. Setup downloads the runtime and a first model: about 2 to 7 GB.
Choose your model and start a conversation.
PAVONA is a Windows beta today. Linux Mint is in validation, and native Apple validation is still pending. About 8 GB of disk space covers the installer, the runtime and a first model of 1 to 5 GB. More models need more room.
A model needs about its own file size in free memory, plus a gigabyte for the conversation. With 8 GB, start with a small model. A graphics card helps but is optional.
Setup measures model fit and reports it as runs well, runs acceptably, runs with limitations or cannot run. Unknown results are labelled.
The installer is unsigned, so Windows may warn once. Check the checksum, then decide. The install guide explains the steps.
New releases
While the app is open, it checks the release list every six hours. Nothing installs until you approve it.
Read the patch notesPatch notes appear in the News panel. “Report a problem” writes a report that you read before it is sent anywhere. Get support.
Benchmarks
Every agent gets the same tasks and the same checkers. The results were recorded on one laptop and are not a promise for yours.
See the resultsMeasured, not claimed
Recorded task outcomes from 2026-09-24. These exploratory results do not establish product superiority: matching model and resource budgets have not been verified. It trails on time per verified task, and the team's standing card is to close that gap and measure again. The whole page, with every task and every weakness.
| Agent | Tasks passed | Seconds per verified task | Files left behind per task |
|---|---|---|---|
| PAVONA | 94% | 43.09 | 0.057 |
| Codex | 94% | 107.147 | 0.371 |
| Claude Code | 91% | 38.914 | 0.114 |
| OpenClaw | 91% | 31.862 | 0.057 |
| Aider | 20% | 63.939 | 0.971 |
| Crush | 83% | 29.013 | 3.143 |
| Goose | 89% | 43.909 | 0.371 |
"Not run: billing" (or auth, model, engine) means the agent's own program refused every task for that reason and nothing was measured. "No verified work" means the agent passed no task, so its speed is not a figure. "Not measured" means the harness cannot observe that figure from outside an external agent.
Benchmark comparison
Select a category, then focus or tap a bar to see its evidence. Highlighted segments show a measured advantage.
Loading the published comparison…
Published results refresh here every minute. A new score appears only after a benchmark has finished and been published. The animation does not mean there is a new measurement.
Check it yourself
The download includes readable source and the test suite.
python run-tests.pySome tests use models, the network or local files. This page does not claim a passing run.
Open the handbookFree while in beta
Your beta copy stays yours for good.
The source is readable and comes under the PAVONA licence. It is not open source.
Now playing in the corner: Feast Your Eyes by Mastodon