Obscura

The problem

Conventional browsers treat profiles as a UX convenience — somewhere to keep your bookmarks separate. They don't treat them as a security boundary. Cookies leak across tabs through clever sites; the browser fingerprint stays roughly constant whether you're in "work" or "personal"; storage isolation is uneven; network egress is shared. For high-trust workflows that's not enough.

Investigators, security teams, journalists, consultants juggling competing client engagements — anyone whose work depends on different identities not contaminating each other — have been improvising solutions: multiple physical machines, VMs, browser-extension hacks. None are good.

The approach

Obscura takes the position that profile = identity, and identity has to switch as a whole. When you pivot to a different profile in Obscura, you don't just get a different cookie jar — you get a different fingerprint, a different user-agent, a different storage partition, optionally a different network egress, and an agentic layer that watches for the kinds of leaks that happen at the seams.

The agentic layer is the part that makes the rest practical. Isolating storage and fingerprints is mechanical; keeping the human from leaking themselves across profiles (pasted clipboard, autofilled email, accidentally-shared screenshot) is the hard part. Obscura's in-browser agents catch the categories of mistake that turn a private workflow into a deanonymizing one.

The build

Obscura is built on a Chromium base with the profile boundary lifted from a UX concern to a runtime concern. Each profile gets its own storage partition, its own network stack with optional per-profile proxy routing, its own fingerprint surface, and its own agent context. Profile pivots happen visibly — the chrome tells you which identity is active so accidental cross-context use becomes obvious.

The agent layer runs locally. No off-device LLM call is required for the routine watchers; the heavier agents (the ones that reason about "does this paste expose something you don't want exposed") can route to a model the user trusts on a network the user controls.

What it does in production

Obscura is the visible artifact of Sophonix's position that AI should run where the data lives, not the other way around. The same constraint shows up in client work: when a project handles sensitive data, the architecture starts from "what can stay local" and earns each external dependency.

Tech stack

• Chromium base with profile-scoped runtime isolation
• Per-profile fingerprint surface and storage partitioning
• Optional per-profile network egress (proxy / VPN / Tor routing)
• In-browser agent runtime, model-pluggable, local-first by default
• Platform: desktop (macOS, Windows, Linux)