Your AI teammate that finishes the job. Takes calls, joins meetings, follows up for days, and asks before it acts. Building in public. npm i @chusky/sdk
We're building Chusky, an AI teammate for small businesses that doesn't stop at a draft.
It takes the call, joins the meeting, and does the follow-up, for days if the job needs it.
Two founders, bootstrapped, building in public. ๐งต
Uber says AI agents booked 4 rides in Halloween week. Airbnb calls agent bookings "very non-material."
The hype is agents shopping for you. The real money is in boring finishing: chasing quotes, confirming jobs, following up.
Where is agent value showing up for you?
@nixonelectrical Useful only if it routes, not just answers. For a trade the difference is capturing the job and pushing a booked slot to your calendar with a confirmation SMS. Anything that can't hand off quotes is just a second voicemail.
@leo_mercerr Day 32 and still honest about $0 MRR - that's the hard part. Meetings aren't the villain, shipping nothing is. Sprint planning beats a post of motivation every time. Keep the cadence.
@Just_Whalen Second question is where the work lives. With Chusky the owner's "yes" is built in: spend, delete, or outbound message pauses and asks first. So a stray instruction past the data rule just hits an approval gate. A lock on the action beats hoping the model holds.
@sunwoopark Yes. Once a fix has a label, matching it becomes a pattern match instead of re-investigating the same gap. The categories we use (intent gap, tool misuse, edge case) make the next deploy apply the fix everywhere, which is what drops the recurrence.
@GANTZZZZZZZZZ Spot on. Kill switches are damage control. The real gate is asking the owner on that exact action. Per-action approval on spend, delete, or outbound is what a small team needs. A red button only cleans up afterwards.
CrowdStrike just shipped a kill switch for rogue AI agents.
A kill switch is a fire alarm. Small businesses need the smoke detector: the agent asks the owner before it sends, spends, or deletes anything.
Would you rather stop an agent or never need to?
Monday, 9am: you can't remember what you promised customers last week.
Fix it today, 10 minutes Friday:
1. List every promise made this week
2. Mark done or not
3. Message anyone waiting, with a new date
Clean slate Monday.
I like the fact that @composio are doing instant but the fact that itโs gated behind the upgrade is terrible, imagine a user with a small agent for only personal tasks and fewer tool calls? Instead of a pay as you go he needs to upgrade? Come onโฆ
AWS just open-sourced a runtime for agents that wait on overnight approvals.
Good. Real work isn't one prompt. It's days of waiting on a human yes. An agent that can't wait patiently and resume is a demo.
Who's building for the waiting?
@HunterLapeyre Most operators miss that agents choke on forms where the phone field normalizes input or the submit is an image. Put a plain phone number next to a real clickable button - turns a missed call into a booking without asking twice.
@tannerchidester Same-day call-back is sharp, but intent cools to zero in 17 minutes. Route flagged chats to a call; if unanswered, SMS the next slot within 20 minutes. Keep callbacks inside the intent window rather than waiting until same day.
@sunwoopark We keep a shared playbook that's versioned. Each fix is categorized (intent gap, tool misuse, edge case), and the matched pattern is applied to every agent on the next deploy. The diff goes to a single source of truth; the agents themselves don't carry separate memories.
@deseventral Identity is the lever. Amazon's block on Muse fits a multi-merchant marketplace because an unidentified agent storing credentials is a liability the host must control. Small-business hedge: a payment flow the merchant owns.
@manish_fp The ask-a-real-customer-a-dumb-question test works: have them call a friend's small business, ask about a price change, and come back with what was said verbatim. People with taste notice the awkward pause you cannot prompt into someone.
@atif2win Add an owner-mid-call scenario. The hard handoff gap is preserving what the agent promised. Chusky asks the owner before changing a booking - same pause that keeps a handoff from spawning new owner work.
Your team nodded through Thursday's meeting. By Monday nobody remembers who owns what.
Fix it before anyone leaves the room:
1. Say each action out loud with one name
2. Add a due date
3. Text the list to everyone within 10 minutes
No name, no date, no action.
@tinycrewgo Clean test. The two misses all three made are exactly what a non-technical owner needs: the concrete PHP version and the "go offline until patched" call. AI flags the topic; the advisory plus a quick host call delivers the action steps.
@Mohitsh71431220 Directories stick when founders return to read submissions about their own tools. Run a weekly roundup where readers nominate the next best submission. Curation beats chasing more listings.
@z3LL@zeebuildshere Agree. A 100-name list beats a broad spray. The reply rate tells you whether the ICP framing or the offer is off. If opens are fine but no one replies, rewrite the first line to name their actual workflow. Positioning comes from patterns in their answers, not a slide deck.