Pagerunner vs Project Mariner
Mariner bet on screenshot-based browser automation as a standalone product. That architecture is slow, expensive, and fragile. The browser-as-standalone-agent thesis is losing to browser-as-infrastructure-for-coding-agents.
Feature comparison
| Feature | Pagerunner | Project Mariner |
|---|---|---|
| Architecture | CDP (fast, reliable) | Screenshot-based |
| Speed | Sub-second | Slow (screenshot round-trips) |
| Privacy | Local only | Content transits Google |
| Price | Free (Apache-2.0) | $249.99/month |
| Bot detection | Stealth mode | Blocked by Cloudflare |
| Unattended operation | Daemon mode | Requires user present |
| MCP integration | Any agent | Gemini only |
| Site intelligence | Learns from usage | None |
| PII anonymization | Local ONNX NER | None |
| Model | Any LLM | Gemini 2.0 only |
Where Pagerunner wins
- Every technical axis. Speed (CDP vs screenshots), privacy (local vs cloud), cost (free vs $250/mo), reliability (DOM vs pixels).
- Bot detection. Stealth mode handles Cloudflare challenges that block screenshot-based extensions.
- Autonomy. Daemon mode runs unattended. Mariner requires a user to be present in the browser.
- LLM flexibility. Works with any MCP client and any LLM. Not locked to a single provider.
- Site intelligence. Adapters and selector tracking compound over time. Mariner starts from zero every session.
Where Mariner wins
- Google ecosystem. Integrated with Gemini 2.0 if you are already in that stack.
- Brand recognition. DeepMind name carries weight in enterprise conversations.
When to use which
Use Pagerunner for
Browser automation. Mariner's architecture (screenshot-based, extension model, cloud-dependent, $250/mo) has fundamental limitations that Pagerunner's architecture avoids entirely.
Consider Mariner if
You're fully invested in Google's AI ecosystem and need a consumer-facing browser assistant (not developer automation).