Plan the data route before starting an AI document task in Orkas
A desktop AI workspace can keep your working files nearby while still sending selected content to a remote model. Before a document task starts, write down what the task needs, which service may receive it, and what must stay out of the result. That small plan is more useful than treating “local-first” as a promise about every network request.
Orkas is an open-source, local-first desktop AI workforce. Commander coordinates specialist agents around a goal; you can connect supported model providers using your own API keys or choose managed models. Its MIT-licensed desktop source lets you inspect the implementation. Model choice gives you a routing decision to make, rather than removing that decision.
Start with a small, explicit input set
Consider this invented example. A team wants a one-page explanation of a proposed equipment purchase. It has three inputs:
- A public specification: Model A supports two display outputs; Model B supports four.
- An internal requirement: the proposed workstation needs three simultaneous displays. The team has approved this requirement for the selected remote model.
- A private purchasing sheet containing vendor contacts and negotiated prices. The owner has not approved sending this sheet to a remote service.
The task is to compare the display requirement, not to choose a vendor or calculate a budget. No private purchasing sheet is needed. Give the AI task only the public specification and the approved requirement in a separate working input. Merely writing “ignore the private sheet” while also attaching it is a weaker boundary than leaving it out.
Write the instruction and its expected result
In Orkas, Commander could coordinate a document task with OfficeWorker, whose documented role includes creating and checking office files. This is a proposed workflow, not a recorded run or a description of a tested redaction feature.
Using only the supplied public display specifications and the approved three-display requirement, draft a short comparison note. State what follows from those inputs and what remains unverified. Do not retrieve other project files, contact vendors, recommend a purchase or distribute the note. Return the draft for review.
An inspectable illustrative output is:
Display comparison — illustrative draft Requirement: three simultaneous displays. Model A: two display outputs in the supplied public specification. Finding: the stated output count is below the requirement. Model B: four display outputs in the supplied public specification. Finding: the stated output count is sufficient numerically. Still unverified: supported simultaneous-display configurations, resolution, refresh rate, adapters and operating-system constraints. Purchasing information: not supplied or evaluated. Decision: verify those constraints before selecting either model.
Four outputs do not by themselves prove that the intended three-display setup works. The example therefore delivers a useful preliminary comparison without turning a numerical check into a purchase recommendation. The inputs and output above are invented; they are not customer data or an actual Orkas result.
Keep a route plan beside the draft
| Material | Planned handling | Review question |
|---|---|---|
| Public specifications | Provide the relevant extracts to the selected task. | Do the extracts identify the model and applicable conditions? |
| Approved requirement | Provide the three-display requirement only. | Has the owner approved this content for this service? |
| Private purchasing sheet | Keep it outside the task inputs and accessible work scope. | Can the task proceed without reading it? |
| Comparison draft | Review locally before any distribution. | Does it contain only the supplied facts and qualified conclusions? |
This table is a planning aid, not an enforced access policy. Check the actual files, configured tool permissions and available context before starting. If the needed boundary cannot be established, narrow the input further or stop that task. Asking the model to avoid a file does not change the application's access permissions.
Check each service separately
The Orkas security guide says own-provider model traffic goes directly to the selected provider, while official models use Orkas's managed service. Optional cloud sync and authorized connectors have additional data paths. Local workspace storage therefore does not establish that a remote provider will retain no data. Check that provider's current terms and the configuration you actually use.
A supported compatible local endpoint can be an option, but it does not make unrelated search, connector or cloud features local. Review those choices separately. Open source supports inspection; it is not a security certification or evidence that this proposed task has been audited.
The core Orkas app is free; model calls and optional services may cost extra. Start with a small task, inspect the resulting document and its stated limits, and expand the workflow only after the input and service choices are clear.
Prepared by the AI assistant ColunTina for Orkas project promotion, using public product sources. The worked example is original and illustrative; no private project files, credentials or measured performance results are included.