Five field leaders. Twenty capabilities.
Sitebridge vs Claude for Chrome, Playwright MCP, Browser MCP, and Chrome DevTools MCP. Every cell traces to a public-facing source or a file path in the Sitebridge repo. No hand-waving. ● ships it ◐ partial ● doesn’t.
| Capability | Sitebridge | ClaudeforChrome | PlaywrightMCP | BrowserMCP | DevToolsMCP |
|---|---|---|---|---|---|
| Bundled with model + agent runtime (no third-party sub) | ● | ● | ● | ● | ● |
| Runs inside your real logged-in browser | ● | ● | ● | ● | ◐1 |
| Works with any LLM provider (no model lock-in) | ● | ●2 | ● | ● | ● |
| Local-only data path possible | ● | ●3 | ● | ● | ● |
| Pre-click safety guards (expectedTarget, viewport check) | ● | ● | ● | ● | ● |
| Undo log + DOM snapshot ring buffer | ● | ● | ● | ● | ● |
| Compound wait-aware tools (click+wait+read) | ● | ◐4 | ◐ | ● | ● |
| Cross-session selector learning | ● | ● | ● | ● | ● |
| Semantic search across open tabs | ● | ● | ● | ● | ● |
| Markdown negotiation (~10× token reduction) | ● | ● | ● | ● | ● |
| Userscript / JS injection / CSS overlay | ● | ●5 | ◐ | ● | ◐ |
| Network capture + custom HTTP requests | ● | ● | ◐ | ● | ● |
| Performance tracing (Core Web Vitals) | ● | ● | ◐ | ● | ● |
| GIF recording with click indicators | ● | ● | ● | ● | ● |
| Element-selection human handoff | ● | ● | ● | ● | ● |
| Domain integrations (Smartling/n8n/Confluence/GDocs/Jira) | ● | ● | ● | ● | ● |
| Per-MCP-client session isolation (Tab Groups) | ● | ● | N/A6 | ● | ● |
| Clipboard-routed secrets (no transcript leak) | ● | ● | ● | ● | ● |
| Customer-held audit + enterprise kill-switch (attestable) | ●7 | ● | ● | ● | ● |
| No category blocklist (finance / high-ROI domains allowed) | ● | ●8 | ● | ● | ● |
| Documented tool count | 30+ | ~5 | 40+ | ~8 | 29 |
- Chrome DevTools MCP can attach to your real Chrome but only after relaunching with --remote-debugging-port=9222 every session — characterized as “torture” in published comparisons.
- Claude for Chrome runs only against Anthropic’s own models, gated to paid plans (Pro = Haiku 4.5; Max/Team/Enterprise = Sonnet 4.5 / Opus 4.5).
- Claude for Chrome inference happens on Anthropic’s servers — fine for many uses, blocking for regulated workloads.
- Implicit browser-native waits, but no documented compound primitive that combines click + navigate + settle + read in one call.
- Anthropic’s blog explicitly lists capabilities as “clicking, form-filling, navigation” only.
- Playwright MCP spawns its own browser — per-client isolation is automatic but at the cost of throwing away cookies, MFA tokens, browser extensions, and your real fingerprint.
- Core edition (SBD-84): hash-chained audit log drains over OTLP to a collector the customer runs; Chrome enterprise policy can kill-switch or deny action classes from the buyer’s side. The vendor cannot override managed policy. None of the four competitors ship a customer-held, attestable control plane.
- Anthropic’s extension blocks financial-services and adult-content sites by default — publicly admitting they don’t trust the agent in the high-value domains where automation has the most ROI. Sitebridge has no category blocklist because the safety stack is structural (audit + undo + pre-click guards), not a list of forbidden URLs.
The numbers and tells behind the matrix.
Pull these when the comparison table alone isn't enough. Each one is single-source and dated.
Finance sites are on their default blocklist.
Claude for Chrome blocks financial-services and adult-content sites out of the box. That is Anthropic publicly admitting they do not trust the agent in the high-value domains where automation has the most ROI. Sitebridge has no category blocklist — the safety stack is structural (audit + undo + pre-click guards), not a list of forbidden URLs.
11.2% post-mitigation prompt-injection rate.
Anthropic's own blog: 23.6% before mitigations, 11.2% after. One in nine attacks still gets through on the official extension. Sitebridge's expectedTarget pre-click guard plus the undo snapshot ring is defense-in-depth for exactly the cases where the model is fooled.
They buy what they can attest.
Regulated buyers do not judge features — they judge controls they can prove. Core edition ships a hash-chained audit log that drains over OTLP to a collector the customer runs, plus a Chrome enterprise-policy kill-switch and mutation gate the vendor cannot override. A bank can prove the control fired in a customer-held, tamper-evident record. None of the four competitors ship this.
One product. Every major model.
Sitebridge routes browser tasks to the model that fits — cheap ones to fast small models, hard reasoning to the big ones. Hover the chart to see how the routing decision lands per task type.
Browser navigation: find-button-and-click. Doesn't need reasoning. Route to the cheapest fast model — that's most of agent traffic.
Try the pre-click guard.
Type what the agent intends to do, then type (or pick) what the page actually shows. The guard accepts, refuses, or asks for help — same logic that gates every click in production.
The pre-click guard compares the agent's stated intent to the dialog actually shown. Hidden swaps, prompt-injected destructive verbs, and ambiguous confirms get caught here — before the click goes through.
Save changes to comment? Your draft will be preserved.