Hark's terms let its assistant read connected email and make authorized purchases

The terms, as TestingCatalog summarizes them, permit transactions when authorized ^[11]. That summary does not describe the authorization step. A confirmation on every purchase and a standing grant at sign-up are different products. Under the first, each charge waits for a person. Under the second, a misread checkout page becomes a charge. Until Hark documents which one it shipped, I would connect one calendar and no saved payment method.

Hark calls the product "personal intelligence" and sets it apart from a traditional chatbot ^[9]. The proactive part requires a minimum level of access. A daily brief that decides on its own whether something needs attention ^[16] has to read the inbox and calendar before anyone asks. It does not work without standing read access to both.

Hark said Handoff topped several browser-use benchmarks ^[17]. Those are results on someone else's sites. Working pages by clicking and typing leaves the agent dependent on sites that change their interfaces, block automated traffic or demand human checks ^[12]. The scores carry over to a user's bank or airline only if those sites hold still and let automation through. The unreleased build adds reusable browser skills, procedures users define for specific websites ^[20]. Per-site skills are a sensible fix for brittle pages. They also move the upkeep of each procedure to the user.

TestingCatalog's captured configuration names five model identifiers across four model families ^[15]. Browser-login settings reference Qwen3.6-27B and Gemini 3.5 Flash Lite, and browser challenge handling names GPT-6 Astra ^[10]. TestingCatalog cautions that these are internal identifiers that do not confirm which model handled any actual request ^[6]. If the split ships as captured, granting account access means trusting the login models as well as the assistant the user talks to.

Hark pitched Handoff at everyday consumer work, demonstrating it ordering gifts and booking appointments ^[4]. Its August preview described linking personal accounts so the agent could draw on saved addresses, preferences and history ^[8]. The personality settings sit in a Markdown "soul" document, so they can at least be diffed ^[19]. The memory controls matter more. If the shipped build lets users inspect and delete stored entries ^[18], I would test that screen first.

Adcock did not say how long the free plan for the first 100,000 sign-ups lasts or what it includes ^[1][2]. According to TestingCatalog, the plan names in development run to four Pro tiers with 5x and 20x usage options, and suggest that higher tiers mainly raise usage limits ^[13].