Original on your diskQ3-report.xlsx09:41 · before the first edit
How it works
How Convira works.
Start with a result you want. Convira gathers only the context you choose, shows the work as it happens, and saves the result on your computer. These six visual chapters explain the experience first - and the technical architecture underneath it.
01Request to result
You describe the result. Convira handles the steps.
A chatbot usually leaves you with instructions. Convira can gather the material, use the right tools, show the work, and save a finished file. Follow one message from send to result.
Describe the result and add any files, screenshots, or folders Convira should use. The source material remains visible before you send.
Q3-report.xlsx · attachedVisible workspace
The answer is not hidden in a progress spinner.
When Convira touches a document, the conversation stays beside the file. You can see cells, pages, or slides change as each step lands, then open the saved result in the app you already use.
02What Convira can see
You choose what Convira can see for this task.
The useful difference is not drag and drop. It is an inspectable context boundary: every source entering the next run is visible, intentional, and removable before you send.
Assembling contextOnly the sources you selected
18% usedUse the right context without opening the whole computer
03Active History
The same session can continue without carrying everything.
Active History separates continuity from accumulation. Keep the decisions that matter, leave stale attempts behind, and prevent on-device material from quietly crossing into a cloud run.
A visible history boundary
Continuity, privacy, and context capacity - controlled separately.
Each runtime keeps its own lane. Include a lane, exclude it without deleting anything, or compact older runs into an inspectable recap.
Board analysis and client decisionsincluded
Payroll details stay on this deviceexcluded
Four runs summarized into one recapcompacted
Continuity without silently carrying stale or private material.
Cloud, Local, Private - every run lands in the lane of the runtime that ran it, with its exact token weight.
Toggle a lane out of the next run, or drag a single run between lanes. What you see is what the model gets.
Dragging on-device history toward Cloud stops the move mid-air and asks. Keep it on device with one click.
Old runs fold into a short recap - the gauge drops from 73% to 30% and the facts stay carried.
04Saved on your device
Sessions, plans, and file backups live on your device.
Convira makes local storage understandable. See how much space your work uses, choose how long to keep it, and restore an original file that was backed up before a session changed it.
Automatic safety copySnapshot 09:41Protected by your retention rules
LOCALVisible working fileQ3-report.xlsx14 edits · saved in place
LiveYour device is the system of recordLocal storage
INSPECTABLEBackups6.3 MB
Sessions24 runs
Plans8 versions
Restore without overwriting the archiveChoose where the recovered copy goes. Your current work and safety copy remain.
RECOVER A COPYThe run record on this device is the authoritative copy. Removing cloud access does not erase the local work you already created.
05Where the AI runs
Choose where the AI processing happens.
Use powerful hosted models, run AI on this computer, or connect a larger machine you control. Flip through the three choices to see the exact path your request takes.
Local runtimePrivate inference that works offline
The model, selected context, and run stay on this computer.Desktop + local model
- Inference happens here
- Works with Wi-Fi off
Managed tool service
- Web, media, connectors
- Never receives the whole run
Tool provider
- Receives that tool input
- Returns that tool result
Durable work stays with youRun record, created files, plans, and automatic backups remain on this device.
Local inference never leaves your computer. The dashed path exists only when you deliberately invoke a managed online tool such as web research, media generation, or a connector.
06Safety and architecture
Convira can act, but it does not receive unlimited access.
Every proposed action is checked against your rules and approvals. Filesystem scope, network policy, and native process isolation provide separate layers, with the exact isolation depending on the host and action.
For technical readers
Four layers, four separate responsibilities.
The interface shows the work, the AI proposes a specific action, the execution layer checks it, and the host applies the available filesystem, network, and process boundaries. No layer gets to silently replace the others.
Chat on the left, live workspace on the right. Every step the agent takes renders here.
A frontier model in the runtime you chose. Its only output is text and typed tool calls - it never touches your files directly.
The execution kernel validates each call against policy, runs it, and journals every write so runs survive crashes.
Filesystem scope and network rules remain separate from native process isolation. Convira applies the layers the host supports and reports the isolation used.
Want the implementation details? Read the full security model →
Now watch it work on yours.
Start with a file you already know and a result you can judge. Convira will show you the work from the first step to the saved result.