Methodology note: This review is based on hands-on testing conducted in June 2026. It includes a full signup through Arlo’s three advertised front doors (Slack, MS Teams, and phone/SMS), a live two-minute onboarding AI voice call, the complete setup wizard, and the dashboard. Arlo is a product of StartClaw and was introduced to me by its founder Idan Mann.
Who It’s For
Arlo introduces itself as a teammate: “basically the teammate who reads your inbox, wrestles your calendar, and handles the stuff that takes ten tabs.” The positioning is a personal/work AI assistant that meets you on whatever channel you already live in, e.g. iMessage/SMS, Slack, or MS Teams, with voice (inbound and outbound) offered as a casual add-on rather than the centerpiece. That alone separates it from the voice-agent-first products this series usually covers: voice here is one channel among several, not the product. But it’s a
The setup wizard’s own questions sketch the intended buyer: team-size options run solo, 2–10, 11–50, 51–250, 250+, and the connectable tools skew toward knowledge work (Gmail, Calendar, Notion, Slack, GitHub, HubSpot, Linear, Jira). It’s an “assistant for busy operators and small teams” play.
Setup Experience
Three entry points are offered: Slack, MS Teams, phone/SMS. The Slack path dead-ended twice. The first attempt completed an OAuth handshake. Arlo found the real workspace, an “arlo” app was approved into it, but then dumped the user back at the channel picker as if nothing happened. The phone/SMS path, by contrast, kicked off cleanly and texted immediately.
The Slack retry was worse, and it produced the strongest receipt in the build. A second Slack attempt threw an “App is not approved by Slack” error alongside a raw database payload surfaced verbatim to the UI:
{"ok":false,"error":"duplicate key value violates unique constraint \"provider_secrets_pkey\""}
Two findings in one. First, there is effectively no error handling at this layer: a Postgres primary-key collision reached the end user untouched. Second, it explains the first dead-end: the earlier broken Slack attempt had already written a secret to Arlo’s database, so the retry collided with it. The first failure wasn’t harmless; it left orphaned state that then blocked the retry. That’s a failure stacked on absent error handling. And it broke the setup flow a second time, forcing a return to arlo.sh to resume.
A few more setup receipts:
- The signup link loops. The onboarding voice call’s entire goal was to text a signup link; tapping it returned the user to the top of setup (re-asking the first name). The working path turned out to be the original browser tab, via a “finish setup” button and not the link the voice agent proudly delivered.
- The integration menu leaks the architecture. Twenty-nine-ish connectable items including Gmail, Google Calendar/Docs/Sheets/Drive/Tasks, Notion, Slack, Outlook, HubSpot, Linear, Jira, Airtable, GitHub, Bitbucket, Discord, Figma, YouTube, Reddit, Twitter — and Composio, Composio Search, Slackbot, Code Interpreter, SerpApi, Firecrawl, Tavily, Perplexity. Those last several aren’t user-facing apps; they’re developer primitives and a third-party integration/auth layer (Composio). The grounded read: Arlo is an orchestration layer riding Composio’s connector catalog rather than building native integrations, and the menu is largely that catalog piped through uncurated. The “Slack” + “Slackbot” and “Composio” + “Composio Search” duplicates are the tell. This is a hypothesis. And, using a well known stack is not any sort of sin.
- Permissions surface is wide. Gmail was connected with read/compose/send/delete scope. Calendar offered an “see/edit/share and permanently delete all the calendars you can access” scope (declined; view/edit granted instead). Meeting settings defaulted to auto-join + record/transcript + auto-share notes (all enabled in this run).
- The “first task” buttons are dead. Step 6 of 6 (“You’re all set. Try a first task”) offered “Triage my inbox and draft replies,” “Summarize what changed overnight,” “Draft the renewal email.” None of the buttons functioned; “Go to dashboard” did.
- Personalization theater. The wizard asked for a first job for Arlo (entered: “research topics for LinkedIn posts”) and then suggested three canned first tasks unrelated to it including “Draft the renewal email,” which has no possible context for a brand-new account. Intent was collected and ignored.
The First 15 Seconds
Two openers, two registers — and the split is deliberate.
SMS opens in a lowercase, millennial-casual voice, emoji and all: “i can also just call you, or pick up if you call me, whatever’s easier.” It even names its own first-impression problem — “tap the contact card i just shared so this thread shows my name and logo instead of a random number. makes the whole thing feel less sketchy lol.” Self-aware spam-avoidance, patched with a vCard and a joke rather than a verified business sender.
Voice opens completely differently: “Hi there. This is arlo, an AI assistant. Who am I speaking with?” The voice level is professional, neutral, no trace of the lowercase persona. Notably, it did not know the caller’s name despite the name having been entered in the SMS flow minutes earlier. Cross-channel identity does not carry.
Call Flow Design
The voice call was very polished. It was goal-oriented (its objective: text the signup link), handled barge-in cleanly, stayed conversational through an off-script curveball, and ended gracefully. Critically, it respected its own channel’s affordances –“Cool if I text you the sign-up link?” – rather than trying to read a URL aloud or referencing something only SMS can do. No inverse-affordance leak.
What They Do Well
- Channel-separated prompts at the content level. The casual SMS register and the professional voice register are genuinely different voices, not one persona piped through two pipes. That’s the higher form of channel awareness, and it’s earned credit.
- Voice craft. Professional delivery, barge-in handling, affordance discipline, graceful close, conversational recovery on a curveball.
- The onboarding-call mechanic. A two-minute voice call to get set up is a nice idea on its own and works well.
Where It Breaks
The cascade, in one place: Slack dead-ends twice (the second leaking a Postgres key); the signup link loops to the top of setup; the “first task” buttons are dead; suggested tasks ignore stated intent; voice doesn’t know your name from SMS; and the integration menu leaks its own plumbing. The throughline is engineering, not craft, e.g. state shared across channels, idempotent retries, error handling, integration depth. Good ideas, but more QA is required.
Who This Is Right For
The core assistant works once you’re past the onboarding. Arlo is for solo operators and small teams who want one assistant across SMS/Slack/voice.
#AI #AI Voice #AI Voice Agent #arlo #arlo.sh #Voice AI