Operate a local Omnigent install — crashes, blank UI, restarts, session attach/resume/dispose, process topology — and drive other agents via the sys_* MCP tools.
Extract a person's LinkedIn writing voice DNA into an 8-dimension style profile with vocabulary fingerprint, signature moves, and a Prompt Engineering Guide.
Audit third-party code (repos, npm packages, binaries, plugins) for telemetry, exfiltration, supply-chain risk, and hostile defaults before install; verdict SAFE/CAUTION/UNSAFE with file:line evidence.
When loaded: deliver reports, intake interviews (>3 fields), and option/mockup choices as one self-contained offline .html page instead of chat Markdown. Forms return answers via a copy-paste token.
Optional layer: write specs, plans, designs, and reports as MDX (Markdown + a component library) rendered locally via Vite — git-diffable, degrades to plain Markdown, no remote services.
Pre-install safety check for a software update — invoke on an update prompt, 'new version available', upgrade notice, or when asked if an installed/trusted tool is safe to update.
Drive a browser via Playwright MCP — navigate, fill forms, click, scrape, or run e2e tests/scripts. Two servers: `playwright` (fresh, no logins) and `user-browser` (the user's live logged-in browser via CDP).
Multi-round multi-source research (Perplexity, Tavily, Exa, Gemini, WebSearch) producing a cited report with confidence scores and gaps.