Show HN: Saccade – Live semantic browser truth for AI agents
Saccade is a browser agent interface that provides AI agents with a semantic, object-addressed view of a single authorized Chrome or Edge tab, eliminating coordinate-based clicking It uses a browser-pushed delta system where agents get an initial full semantic snapshot, then receive only changes (deltas) after a known revision, avoiding full-page re-reads Actions include local verification of visibility, enabled state, stable geometry, and authority before dispatch, with receipts confirming sema
Analysis
TL;DR
- Saccade is a browser agent interface that provides AI agents with a semantic, object-addressed view of a single authorized Chrome or Edge tab, eliminating coordinate-based clicking
- It uses a browser-pushed delta system where agents get an initial full semantic snapshot, then receive only changes (deltas) after a known revision, avoiding full-page re-reads
- Actions include local verification of visibility, enabled state, stable geometry, and authority before dispatch, with receipts confirming semantic transitions
- The 0.2.0 release candidate achieved 24/24 ordinary targets and 24/24 canvas reflex targets with zero misses across Chrome and Edge, with mean action latency of 7.22–8.50 ms
- Saccade is designed for live agent work in authorized user sessions (admin pages, forms, iframes, uploads) rather than scripted testing, differentiating it from Playwright
Why It Matters
Saccade addresses a critical gap in browser automation for AI agents: the inefficiency and fragility of screenshot-based or coordinate-based interaction. By providing semantic object identity and browser-pushed state deltas, it reduces token consumption and action latency while improving reliability on complex, dynamic pages. This approach is directly relevant to anyone building production browser agents that need to interact with real-world, signed-in web applications.
Technical Details
- Tab lease model: Each request names a
tab_id; one tab has one writer and stays isolated from other agent sessions, preventing concurrent modification conflicts - Semantic truth system: Agents read structured semantic objects (text fields, checkboxes, buttons, iframes, tables, etc.) rather than raw HTML or screenshots; the browser pushes revision deltas after an initial bounded full view
- Local action verification: Before dispatching any action, Saccade checks visibility, enabled state, stable geometry, topmost state, and current action authority under one deadline, returning a receipt with the resulting semantic transition
- Form optimization: Independent form fields can be preflighted and sent as one batch; submit, navigation, and upload remain explicit actions, reducing model calls
- Benchmark performance: In controlled same-model form comparisons, Saccade used 8 browser tool calls vs. Playwright's 10 and spent 1.52s vs. 6.88s inside the browser MCP path; total task time was 38.78s vs. Playwright's 26.20s
- Supported semantic objects: Text entry (fields, search, areas, contenteditable, iframe editors), choice (checkboxes, radios, switches, selects, ARIA listboxes), navigation (buttons, links, tabs), files (standard inputs, upload triggers), page structure (headings, paragraphs, lists, tables, alerts, images), and composition (same-origin iframes, open shadow roots)
- Requirements: Node.js 18+, Saccade Extension in Chrome or Edge, installed via
npx -y @nanlogic/saccade installwith MCP server atnpx -y @nanlogic/saccade mcp
Industry Insight
- The object-addressed, semantic approach represents a shift away from vision-based browser automation toward structured interaction models—teams building browser agents should evaluate whether semantic tooling reduces their error rates and token costs compared to screenshot-based approaches
- Saccade's session-isolation and lease model solves a real production problem: concurrent agent sessions corrupting each other's state—this pattern should be adopted by any multi-agent browser automation system
- The benchmark results show a trade-off worth considering: Saccade reduces per-action overhead and tool calls but may have higher total task latency on simple workflows; the value proposition strengthens on complex, multi-step forms and dynamic pages where Playwright-style locators fragility increases
Disclaimer: The above content is generated by AI and is for reference only.