01
Web page
The user opens a real page with its content and interface.
SIH 2026 / Problem ID SIH26171
Let AI see.
Keep data private.
On-device visual perception for lightweight browser agents. MaskAgent understands a web page, masks sensitive context, and acts locally through Ollama.
A privacy layer
before intelligence.
Traditional browser agents pass raw page context to an AI service. MaskAgent inserts a deliberate boundary: understand the interface, remove the secrets, then let a local model decide.
02
Scan UI
DOM structure and visible elements are mapped for context.
03
Detect PII
Names, email, phone, address and password fields are identified.
04
Redact
Private values become safe placeholders before inference.
05
Act locally
Ollama chooses a browser action from the sanitized context.
Small agent.
Serious guardrails.
Built for useful browser automation without unrestricted access to personal information. Every action stays legible to the person using it.
Privacy-preserving processing
Redacts sensitive UI values before they enter an AI prompt.
DOM + visual understanding
Combines page structure with visible UI signals to find the right target.
Local browser actions
Supports CLICK, TYPE, SCROLL, SELECT and DONE through a local model.
Human-readable activity
An activity stream shows scans, redaction, decisions and execution.
Action validation
Checks whether a target exists and an intended action is relevant.
Research-ready foundation
Designed to grow toward OCR, prompt-injection defense and cross-browser support.
CONCEPT &
WORKING DEMO
A working prototype is easiest to understand when the privacy boundary is visible. Watch both videos, then explore the interface captures below.
The people behind the prototype
Built for the next web.
MaskAgent is a Smart India Hackathon 2026 project exploring how browser agents can be capable and privacy-conscious at the same time.
“Private information should be protected before AI processing, not after it.”
Technology stack
- Interface: HTML, CSS, JavaScript, Chrome Extension APIs
- Inference: Ollama with DeepSeek-Coder
- Browser: Chrome, Brave and Chromium-compatible browsers
What comes next
- OCR-based visual PII detection
- Prompt-injection protection and permissioned actions
- Structured decisions, better planning and Firefox support