Miles Matthias

onboarding cli flow is nice, well done

running on my M4 MBA, no performance issues so far

"channels" should be top priority to configure E2E and have working asap, as it's the front door

chatting via a personal slack only (imessage started randomly responding to people with claim codes, which wasn't obvious to me that it would happen)

it being able to actually do things is the selling point, which it has many integrations, but certainly running into the limits of the nascent agentic economy -- oauth times out, 1password/keychain require running back to the MBA to do TouchID/password, google account for clawdbot got shut down, etc.

my personal goal is to have it make phone calls for me via , , funnel. despite seeing loads of people on X saying that it made phone calls for them, it's still not working for me after hours of debugging

one successful use case: it's the only AI bot that has been able to check my email every day, update a gsheet based on emails received, and then give me an updated output from the sheet (it's a personal finance use case leveraging data imported by ). ChatGPT & Claude recommended writing a Google Apps Script, Gemini couldn't string multiple gsuite app use cases in a row

there's also something odd going on with sessions when talking to openclaw via a dedicated channel, forgetting context and capabilities. I gave it access to my opentable account to make a dinner reservation, saw it open chrome to log in and check availability. Then later I asked it to make some reservations and it said that it did. But nothing showed up in my account...

odd to me that sessions inherit model context windows, just let me talk and you organize the context length as needed. I never hit these errors using LLM chats directly, nor do I hit these errors on the clawdbot tui, only when chatting via a channel (ex: slack)

would love to see more opinionated user flows and setups make their way to the docs/onboarding guide. It'd be cool to see them even be generated dynamically. Onboarding could ask "what's the number 1 thing you want to do with this tool?" and it configure a solid, well-tested route. Or it could default to different modes -- chatbot, developer, etc.

the security implications are real and scary, be cautious. I have yet to explore the marketplace of openclaw skills as we don't yet have great LLM injection protections.

the success of the project has proven one thing -- users (technical and non-technical alike) are hungry for AI agents to actually do stuff and have impact in their lives

"connecting to xyz" announcements are good, but agents will win when they can show real problems solved automatically

there's an entire economy of services to be built specifically for agents to get things done, along with protocols to be defined on how agents coordinate with other agents and services