What can Osaurus AI agents do?

Osaurus agents run on your own computer rather than in a vendor cloud, and the tasks shown on the site are document-centric: reading files you point them at, reasoning across several of them, and writing results back to disk. If your work involves local documents you'd rather not upload, that's the fit. If you need web-connected or team-collaboration features, nothing in the available material describes those.

The two task patterns shown

1. Review and edit a single document

The site's contract example walks through a full edit loop:

• Input: a local file (MSA-v3.docx) plus a rules file (Playbook.md).
• Action: the agent reads both, compares the contract clause against the playbook, and identifies a conflict — §9's uncapped liability versus the playbook's requirement that liability be capped.
• Output: a proposed edit to the document itself, shown as a diff (− Liability is uncapped for all claims. / + Liability is capped at twelve months of fees.), which you then Approve or Deny.
• Result: the change is marked "Approved · tracked in Word," with an Undo option.

The visible reasoning step ("I'll read §9 in MSA-v3.docx, then check it against the liability rules in Playbook.md") matters because it tells you what the agent is about to touch before it touches it.

2. Work across a folder of files

The second example is a summarization job over a directory:

• Input: a folder (~/Practice/Intake) containing intake-0914.pdf, session-notes.md, referral-letter.docx, and handoff-summary.docx.
• Action: the agent reads all three source files and extracts symptoms, medication changes, and follow-ups.
• Output: a new file, handoff-summary.md, with counts (3 patients, 1 medication change, 2 follow-ups due Friday).

The pattern to notice: the agent decides which files in the folder are relevant, then produces a new artifact rather than editing in place.

Helpers shown in the interface

The app organizes work into agents and projects. The examples on the page include:

The sidebar also shows a Projects structure (Practice/Intake), so files and helpers appear to be grouped by project rather than run ad hoc.

What the evidence does not cover

• Model details beyond naming. The page references "Raptor 4B" and "Raptor 0.6" but gives no capability benchmarks or context limits.
• Pricing. A pricing page exists at osaurus.ai/pricing, but no figures appear in the available material — check it directly rather than assuming a free tier.
• Platforms. Download buttons for Mac and Windows are shown, with Windows labeled Beta. No Linux option is mentioned.
• Anything outside local files. No email, calendar, browser, or API integrations are described.

Deciding whether to try it

Osaurus is worth testing if your task looks like the two examples: a bounded set of local files, a rule or reference document to check against, and an output you can review before accepting. The approve/deny step and the visible file list give you a checkpoint that a cloud agent sending your documents off-machine wouldn't.

It's a weaker fit if you need the agent to reach beyond your machine, or if you need to know cost and model limits before installing — those require a look at the pricing page and the app itself, since the site doesn't state them.