Architecture

Jumper is designed as a browser-native bridge with clear trust boundaries.

Mental model

1. The user is authenticated in a SaaS tab.
2. Jumper maps visible UI elements into structured fields.
3. A local or controlled MCP endpoint receives normalized payloads.
4. AI tools consume those payloads through MCP contracts.

Runtime components

• Browser tab context: source of truth for visible data.
• Jumper extension core: mapping, transform, policy, and audit controls.
• Console and side panel: operator feedback, mapping management, replay.
• MCP bridge endpoint: tool-facing interface for downstream systems.

Data flow

1. Selector resolves against the live DOM.
2. Raw value is extracted as text or attribute.
3. Transformer normalizes to a typed schema.
4. Payload is sent to MCP bridge.
5. MCP tools execute downstream actions.

Security posture

• Session-aware extraction occurs in the authenticated browser context.
• Least-privilege mapping limits exposed fields to explicit selectors.
• Audit visibility keeps bridge operations reviewable.
• Local-first bridge option reduces external data propagation.

Operational recommendations

• Treat mapping configs as code and version them.
• Prefer stable selectors anchored to semantic attributes.
• Gate production actions behind confirmation policies.
• Monitor payload drift when SaaS UI versions change.

Next: use Documentation hub and MCP registry to scale beyond one-off mappings.