Crew is real and it acts - it chats in your language, remembers you, browses the web, drafts real Gmail, runs jobs while you sleep, and greets you with a recap. The core "not-just-ChatGPT" engine is live in production.
The gap is the last 10% that makes people believe it: too much good work is built but switched OFF, sitting on a staging branch, or reachable only if you know the magic words. A normal person does not yet reliably feel the "it did a whole thing for me" moment on their first visit.
Biggest single gap The best work is stranded. 28 commits of UX fixes are on staging and not on prod; the job-search wedge and WhatsApp are built but hard to reach; a security fix is frozen. Finishing and shipping beats building more.
Areas · most-unfinished first
Autonomy · works while you're away
55%
The engine that does work on its own runs, but the "whoa, it did a real thing overnight" moment is not yet reliable or loud for a first-time user.
Done
Per-user scheduler fires every 2 min (proven)
Background worker does long tasks, comes back "on it, ~10 min"
Village home + "while you were away" recap
Push notifications work on device
Unfinished
Village = real live work, not a generic hi/bye loop
Cloud browser that clicks/acts is built but OFF
The overnight "receipt with your name + hours saved" bet not proven end-to-end
Auto-cook backlog queue stalled (17 tickets)
Not ChatGPT: this IS the differentiator - work happening without you asking. It exists; it needs to be visible and trustworthy, not theatre.
Connectors · Gmail & your apps
55%
Crew can genuinely touch your accounts - draft a real Gmail, add a calendar event - but most of the "act on my connected app" story is safety-gated or only half-surfaced.
Done
Real Gmail draft + Calendar event (proven on prod)
Searchable app catalog with real logos
Google consolidated to one connect
Disconnect / revoke works
Unfinished
Reading your inbox needs a paid Google review + your go
Ad/social platforms need a human-approval tap (ban risk)
A security fix (email-to-connector injection) is frozen pending review
Partial-connect state can hide silently
Not ChatGPT: acting inside your real accounts is core. Drafting works; the frozen safety fix must clear before we widen write access.
Jobs use-case · the first tester's goal
60%
The strongest wedge. Crew finds real jobs, reads your CV, and now gives a blunt "why you get ignored" diagnosis from chat - but the auto-apply and interview-watch half is off, and there's no dedicated jobs surface.
Done
Real listings (Indeed via crew-jobspy)
CV-first gate; jobs auto-file into a silent tracker
Positioning/diagnosis coach now callable from chat (was dark)
Apply-prep drafts a filled application
Unfinished
Batch auto-apply built but OFF (needs approval tap, ban risk)
Interview watch (inbox monitor) OFF
No dedicated jobs board / apply-board UI
Job-swipe UI is a mockup only
Not ChatGPT: ChatGPT gives advice; Crew finds real postings, diagnoses your CV against them, and prepares the application. That's the demo to nail.
Onboarding · the first 60 seconds
60%
A better onboarding (sign-in first, honest real first-win, skippable tour) is built - but it's flag-gated OFF and not on prod, so new users still hit the old flow that an earlier reviewer scored 6.5/10.
Done
Onboarding v2 built (welcome, name+goal chat, honest first-win)
Conversational, goal-first
Sign-in first (no login wall at the end)
Unfinished
Flag OFF - not the default, not on prod
Needs a real-device walk in the user's language
Needs a QA pass before promote
"Get started" and empty-name edge cases
Not ChatGPT: the honest first-win (Crew does one real thing) is the anti-chatbot moment. It's built - it just isn't turned on.
Memory · knows you
65%
Crew genuinely remembers facts about you across sessions and self-corrects them. The deeper "builds a living profile of how you like things" engine is built but switched OFF.
Done
Per-user memory (Supabase + pgvector), live
Self-correcting: adds/removes facts as they change
Nightly memory consolidation in the worker
Unfinished
Auto-persona engine built but OFF (no DB reads yet)
Significance scoring specced/half, not wired
Memory-at-scale is a spec
"Paste your AI profile" onboarding step not shipped
Not ChatGPT: continuous memory-of-you is the personalization moat. The base is real; the richer profile engine is dark.
Mini-apps · build-an-app for me
65%
Crew builds real, working little apps from tested building blocks and fills them with your data - a genuine "it made me a tool" moment. Depth and reliability on heavy builds are still rough.
Done
Block engine live (counter, log-form, stat-ring, entry-list)
Agent seeds real rows into the app
Full HTML apps with real interactions, delivered as cards
Unfinished
Heavy builds can race the timeout / show jargon
Empty log-form / edit submits die silently
A one-shot build can wrongly become a daily job (bug)
Creator/marketplace economy is paper only
Not ChatGPT: Crew doesn't tell you to build a tracker - it builds and fills one. Strong; needs reliability polish.
Money / billing
30%
Cannot take money yet. Cost metering is honest and live, but the paywall is intentionally not wired - revenue stays a true $0.
Done
Per-user cost metering, live "Money" view
Daily caps (voice, image, global)
Free-tier model defined (3 topics)
Unfinished
Checkout shows "billing coming soon"
Payment persistence deliberately NOT built
Live purchase never tested on prod
Server-side Stop is a no-op (some stopped runs still billed)
Not ChatGPT: billing is table-stakes, not a differentiator - fine to leave last, but it blocks charging anyone.
Chat · the base layer
90%
Solid. The core conversation - multilingual, streaming, with web browsing and tool use - works well and is the foundation everything else sits on.
Not ChatGPT: this part IS chatgpt-like - and that's fine. It's the plumbing, not the pitch.
Topics · your life-areas
65%
Topics exist as ongoing folders for big life-areas. The decision to make a tapped topic open into a continuable chat (like Claude) is set but not fully built.
Done
Topics as managed life-areas (not a dump)
Auto-classify: is this even a topic?
Per-topic goal evolves from chat
Unfinished
Tapping a topic should open a continuable chat, not a wall of text
Clear Topic vs Project separation still requested
Chat-history view alongside topics
Not ChatGPT: a living thread with status/progress per life-area beats a flat chat log - partially there.
Notifications
70%
Push works end-to-end on device. The settings screen to control them still says "coming soon".
Done
APNs push proven on device
Return-moment recap + bell
Comebacks file into their topic
Unfinished
"Notification settings · coming soon" stub still shipped
Web-push path half-built (no subscribe UI)
Not ChatGPT: the outbound "greet you where you live" channel is what makes autonomy land - the pipe works, controls are missing.
iOS / TestFlight
70%
A real iOS build ships to TestFlight. App Store review is blocked on one small merge (a review-only sign-in bypass).
Done
Capacitor wrap, builds upload (build valid on ASC)
Account deletion + connector revoke (Apple 5.1.1)
AI-processing disclosure on welcome/auth
Unfinished
App Review blocked - needs the OTP-bypass merge (already reviewed)
Native Swift app is a standing track, not done
Privacy/Terms drafts diverge from shipped copy
Not ChatGPT: distribution, not differentiation - but it gates getting Crew into real hands.
Multilingual
85%
Strong. Conversation mirrors the user's language and the UI auto-localizes; a few raw untranslated keys still leak through.
Done
Server-side language mirroring, all agents
First-run UI auto-localizes to device language
French crash + auth bugs fixed
Unfinished
Occasional raw keys (e.g. WA_NOT_READY) shown to users
Full-app i18n coverage still in progress
Not ChatGPT: mostly table-stakes, done well - matters for the SEA/MY audience.
Differentiation vs ChatGPT · the whole point
55%
Every ingredient that makes Crew NOT ChatGPT exists in code - acting, autonomy, memory. The unfinished job is making a normal person FEEL it in one session without knowing any secret flags.
Done
Acts (Gmail/Calendar), browses, runs jobs, remembers you
Works overnight + greets you with a recap
Builds real mini-tools filled with your data
Unfinished
The best proof is OFF or on staging, not on prod
The "receipt with hours saved" moment unproven for a fresh user
Landing still an exact copy of cofounder.co
Not ChatGPT: this is the scorecard for all the above. The engine is built; the belief is the last mile.
Never finished · tick what goes in
FINISH = worth completing · MERGE = just ship it · DECIDE = commit or kill · KILL = delete dead weight
Promote the staging UX batch MERGE
What it takes: you test staging, then one promote. 28 commits of real fixes + honest states are stranded off prod.
Merge the App Store review bypass MERGE
What it takes: merge one already-reviewed branch. App Review is blocked without it.
Unblock the frozen security fix MERGE
What it takes: Edward signs off the email-to-connector injection fix, then deploy. It's a real vulnerability, currently frozen.
Turn on the better onboarding FINISH
What it takes: real-device walk in the user's language + QA pass, then flip the flag on. Already built.
A real jobs surface (board + apply) FINISH
What it takes: one screen for the diagnosis + apply-board. The engine and diagnosis coach already work.
WhatsApp front door FINISH
What it takes: a connect-code screen + a working test number (Meta app currently blocked). Backend is live-dark.
Cloud browser that clicks/acts DECIDE
What it takes: commit to a substrate + one demo flow, or officially move it to V2 and stop half-carrying it. Built, flag OFF.
Auto-persona / knows-you engine DECIDE
What it takes: turn it on behind measurement, or park it. Currently built but reads nothing.
Batch auto-apply + interview watch DECIDE
What it takes: a human-in-loop approval tap (ban risk) + a UI, or park to V2. Server side built, OFF.
Standing goal (invisible plumbing) DECIDE
What it takes: land the good commits or kill the branch. It's a stated north-star with 0 references in prod.
Billing / take real money DECIDE
What it takes: wire Paddle persistence + one live test purchase. Deliberately $0 today.
Kill dead endpoints + stubs KILL
What it takes: delete noor.js, legacy voice-token, native Gmail OAuth, web-push; remove "coming soon" menu items. Pure cleanup.
Fix + restart auto-cook queue FINISH
What it takes: fix the one-shot-becomes-daily bug, then triage/kill the 17 stalled tickets.
Own landing page FINISH
What it takes: replace the cofounder.co copy with our own hero + a real "replay" of a crew session. Still a clone today.