The blocking ask covers most of what you would want and returns the finished answer rather than a stream to babysit; poll and stop exist for the agentic runs long enough to wander, which is a real failure mode when the model is browsing on your behalf. Mode is the lever most worth using deliberately — research and labs behave differently enough from plain search that setting it belongs in the same turn as the question.
A server that drives the Comet browser, sending prompts to Perplexity inside it and returning the completed response.
- A prompt sent and the complete response waited for, optionally in a new chat, with a timeout you set
- Agentic runs monitored through a status poll and stopped mid-flight when they go off track
- The Perplexity mode switched between search, research, labs and learn, or read back when you are unsure which is active
- A screenshot of the current page
- The browser connection established on demand, starting Comet if it is not already running
Comet installed on the machine, since comet_connect starts it and every other tool acts inside it, and a Perplexity session in that browser — routing through a real browser is what makes login walls and dynamic pages reachable in the first place.
One command — npx -y comet-mcp