Numbers worth memorizing
Six facts that close the room.
Each one is single-source, public-facing, and dated. Quote them in any order.
0.0%
Anthropic’s own post-mitigation prompt-injection success rate (down from 23.6%). One in nine attacks still succeeds. SiteBridge’s expectedTarget guard plus undo snapshot is defense-in-depth for exactly the cases where the model is fooled.
claude.com/blog/claude-for-chrome
114k → 27k
Tokens per task: Playwright MCP vs Microsoft’s own @playwright/cli. A 4× tax for using MCP at all. SiteBridge gets to the 27k-class budget via compound tools + markdown negotiation without abandoning MCP.
playwright.dev/mcp · Microsoft 2026 docs
0+
SiteBridge tools shipping today vs Anthropic’s ~5 documented primitives, Browser MCP’s ~8, DevTools MCP’s 29 (debugging-focused), Playwright’s 40+ (sterile-browser). Surface area is not a marketing number when every tool encodes a learned workflow.
app/chrome-extension/entrypoints/background/tools/
500 + 50
Action log ring buffer (500 mutating clicks) plus pre-click DOM snapshot ring (50). After-the-fact “what did the agent do?” is an answerable question. Nobody else ships this primitive.
docs/ARCHITECTURE.md · safety-guards.ts
0
Domain integrations baked in: Smartling job-sweep, n8n node configuration, Confluence whiteboard read, Google Docs replace via official API, Jira form workarounds. None of the four competitors ship any.
background/tools/browser/{smartling,n8n,confluence,google-docs,jira}-*.ts
~10×
Token reduction on documentation sites via markdown negotiation — probe ${url}.md and Accept: text/markdown before falling back to HTML. Not glamorous, compounds across every doc-heavy session.
background/tools/browser/get-web-content.ts
How to use these
One number per moment, not all six at once.
- Hero claim follow-up: 11.2% (the prompt-injection watermark).
- Cost objection: 114k → 27k (token efficiency story) plus our LiteLLM cost-routing.
- “Why not just use Claude for Chrome?” 30+ (surface area) and 500+50 (safety primitives).
- Smartling/CAT pitch: 5 (baked integrations) and the demo of smartling_apply_batch.
- Token-cost-conscious buyer: ~10× (markdown negotiation) and complexity-routed inference.