Diali: An Always-On AI Assistant That Lives in WhatsApp or Telegram

Diali's starting point is simple: an AI assistant should show up where a person already spends their day, not live behind another login in another tab. Once set up, the assistant answers inside WhatsApp, Telegram, Slack, Discord, and several other channels, and a task gets handed to it the same way a message would go to a colleague. Under the hood, each assistant runs the open-source OpenClaw or Hermes engine inside its own private browser, with an optional Linux desktop available as an add-on for heavier work. Either surface can be watched live as the assistant works through a task, and a user can take over mid-task if something needs a human hand. None of this depends on a laptop staying open or a home server staying online: the assistant keeps running on Diali's own infrastructure. Running OpenClaw or Hermes elsewhere usually means picking a tradeoff: on a personal machine, the assistant goes offline the moment the machine sleeps; on a rented server, it stays up but turns into an ongoing maintenance job; on a generic one-click agent host, setup is handled but the AI model is often billed separately and the isolation behind it has to be taken on trust. Diali's own infrastructure is built to sidestep all three: always-on by default, isolated per assistant, with AI billing left to the user's choice rather than folded into a markup.

Diali hosts two distinct open-source engines rather than building its own, and a user picks the one that fits the job, with the option to run one of each on the same account. OpenClaw is the task-oriented engine: it works through multi-step jobs on its own, operates a user's own connected devices, and is the only one of the two that can place phone calls. Hermes leans conversational, built around memory and proactive follow-up rather than task execution. The two also diverge in the details that matter once an assistant is live. Browser automation works differently on each, scripted on Hermes versus a fuller tool set on OpenClaw, and the approval flow for sensitive actions differs too: OpenClaw exposes an Approvals tab with switches for sensitive actions, such as messaging other people or running code, while Hermes uses an approval mode with a timeout. AI provider support splits as well, with DeepSeek available on Hermes only and Mistral, Groq, Chutes, and Ollama available on OpenClaw only, so the right engine can depend on which model a user wants to run.

Rather than locking usage to one billing path, Diali offers three ways to cover the AI model behind an assistant. The default is Diali's own included credits, billed as part of the plan. A user who already has a preferred provider can bring their own API key instead, across ten supported providers, and pay that provider directly. The third option is signing in with an AI plan a user already holds, such as an existing ChatGPT, SuperGrok/X Premium+, or MiniMax subscription, so a subscription that would otherwise sit unused covers the assistant's usage too. That flexibility extends to the model itself: Diali's own credits cover Claude or GPT, chosen per assistant, while a bring-your-own-key connection opens the door to Gemini, Grok, DeepSeek, Mistral, Groq, and others, which matters once OpenClaw and Hermes are being used for different kinds of work on the same account.

Every assistant runs alone inside its own gVisor sandbox, on infrastructure based in the EU (Google Cloud in Belgium, with some services on AWS in Ireland). Secrets and API keys are encrypted with AES-256 in Google Secret Manager, passwords are stored as argon2id hashes, and traffic runs over TLS behind Cloudflare, with upstream OpenClaw and Hermes updates tested before they reach a live assistant. On the data side, Diali states it doesn't read conversations or train models on user data, and messaging-app chats stay inside the assistant's own workspace rather than Diali's. Deleting a thread removes it from history, and closing an account erases what's stored. Support staff can only open an assistant's stored data with the account holder's written sign-off on a support ticket, which keeps the sandboxing principle consistent with how the account is handled day to day.

The clearest fit is an individual who wants an agent available around the clock, reachable through chat apps already in daily use, without taking on server maintenance or device uptime as a side project. Starting at $29.99 a month, the Starter plan's two-hour check-in cadence (for OpenClaw assistants) suits lighter, personal use, while Pro and Max step up to hourly and thirty-minute check-ins with proportionally more included AI credits, a better match for someone running an assistant against more frequent or time-sensitive tasks. The 7-day Starter trial is enough time to connect a real channel, hand the assistant an actual task, and judge both engines' feel before deciding which one, and which plan's check-in frequency, fits the way the assistant will actually get used.