nicoolAI is getting better at working in the background: reminders and recurring tasks are easier to manage, notifications bring you back when an answer is ready, and conversations hold onto more of your work.
Create a simple reminder just by asking. Reminders and recurring tasks now live together in one schedules view, where you can review and manage everything nicoolAI is set to do later.
Get an “answer ready” notification on the web or your phone when longer work finishes, then tap it to jump straight back into the conversation.
Unsent messages are now saved in your browser, so you can move between conversations without losing a draft.
Drop files anywhere onto a chat to attach them. Larger uploads are more reliable across web and connected channels, with clearer progress while a file is being prepared.
Sources from web research now appear as clear, clickable citations inside answers.
Mobile conversations now load earlier messages and show more of your recent chats. Sign-in also survives app updates more reliably, and model selection is steadier while choices are loading.
Slack conversations can now use the relevant Google Drive links and files shared in the thread, giving nicoolAI the context it needs without making you repeat yourself.
Chats feel faster and calmer, with quicker thread switching, immediate feedback while nicoolAI starts thinking, cleaner spacing, and horizontally scrollable code blocks.
Keep talking. I’m listening.
Conversations feel more natural, memory is easier to understand, and your assistant has new ways to work with the information that matters to you.
Send another message while nicoolAI is still working. Your follow-up joins the same conversation naturally instead of starting a separate reply.
Telegram conversations now stay in sync with nicoolAI, including group chats, live replies, and messages sent by your automations.
When nicoolAI remembers something new, you’ll see a small “memory saved” note beneath the conversation. The redesigned memory page also makes saved details easier to review.
New data workspaces give nicoolAI a safe, organized place to work with lists, records, and other structured information on your behalf.
Explore the new skills catalog to see what your assistant can do and which capabilities are available to you.
The mobile app now makes full use of the iPad screen, while cleaner code blocks and a new glossary make chats and docs easier to read.
Channels, one conversation core.
The channel foundation is moving onto one canonical conversation model, with Chat SDK handling native platform transport while Convex remains the durable source of truth.
Chat SDK and the enabled Discord, Slack, and Telegram adapters moved together from 4.32 to 4.33; WhatsApp remains parked until its Cloud API setup is ready.
Channel work now targets native threads, typing, reactions, streaming edits, attachments, and automation replies without making transport state authoritative.
Choose your engine.
nicoolAI can now run a conversation through either a ChatGPT or Claude subscription. The provider layer, model picker, and supporting runtime all learned how to keep that choice explicit, stable, and safe across web, mobile, and background work.
Connect Claude Pro or Max from settings, choose Claude models in web or mobile, and keep each conversation pinned to its original provider so its context never jumps engines mid-thread.
The Codex model list now comes from the live catalog instead of a hard-coded snapshot, with unavailable saved preferences cleaned up automatically.
Conversation pins landed on web and mobile, making important threads easier to keep at the top of the list.
A new command palette lets you search visible conversation threads and jump between them without leaving the keyboard.
MCP setup now detects and fills in server details automatically, while sanitizing discovered metadata before it reaches the form.
GitHub setup now reports the app's live status and gives clear feedback throughout the OAuth connection flow.
Shared-channel memory is now owned by the room rather than an individual participant, and private user-fact writes are suppressed in channel conversations.
Mobile sessions recover from transient token-refresh failures without signing you out, while sign-out and concurrent refreshes are fenced against races.
Runtime hardening covered expired attachment links, inbound-tool audience checks, MCP OAuth failures, Claude web search, and subscription-aware billing for auxiliary model calls.
Consolidation day.
The day after the sprint: fewer new toys, more load-bearing structure. Everything the agent executes now runs in one sandbox, and MCP servers went from plumbing to a user-facing feature.
One execution environment: the legacy bash runner is gone — every bash root now runs in the Vercel Sandbox.
Google Drive is materialized into the sandbox at /gdrive with per-caller authorization, so the agent works on your files like local ones.
Scheduled automations became Convex-triggered, Vercel-executed cron runs — same sandbox, same policy checks.
MCP servers went live in the product: a first-class /dashboard/mcp section, and connected servers now work on inbound channels (WhatsApp, Slack, email) too.
The pack install engine grew a reconciliation core — plan, apply, and export a tenant back to a pack — hardened through four review rounds.
Chat polish: a user dropdown in the sidebar footer with a real profile widget, and a duplicate-persistence fix.
one working day
The 127-commit day.
From 10:00 on July 7 to 01:47 the next morning, eight workstreams moved in parallel — two Claude subscriptions and one Codex subscription running agents against the same repo, with every change going through review before it landed. Here is what actually shipped, in more fidelity than a tweet allows.
01
The chat runtime moved onto Convex
convexstreamingauth
Every web chat turn now runs end to end as a Convex httpAction — streaming, tools, memory, file uploads, Google Drive mounts, and the sandbox bridge all execute next to the data instead of inside Next.js. After the two runtimes reached parity, the old path and its CHAT_RUNTIME feature flag were deleted.
Chat requests carry a JWT that Convex verifies itself — an invalid token is now a clean 401, not a 500.
Resumable streams survive refreshes without phantom replays, stop polling on cancellation, and garbage-collect their chunks after every turn.
Conversation titles, reasoning traces, and memory writes persist from the same action that streams the reply.
02
Sandbox v2 and a real policy layer
sandboxauthzbilling
The execution sandbox upgraded to @vercel/sandbox v2 with a native keep-alive lifecycle, and every piece of v1 workaround machinery was deleted. Underneath it landed a new policy core that decides what any given turn is allowed to do — and who pays for it.
Work units with deterministic keys and scope shares define what a turn may touch; revoked and expired shares filter out by default.
A single operationAuthz choke point now fronts every operation, with adapters per surface.
Inference funding is enforced per turn, so usage is always attributable to a plan that covers it.
Four review waves hardened lock races, expiry-bounded renewals, and snapshot-commit status checks.
03
Bring your own MCP servers
integrationsmcp
You can now connect your own MCP servers to your agent: an encrypted credential vault, OAuth flows for servers that need them, and a dashboard section to manage everything. A new run_integrations_code tool gives the agent code-mode over every server you connect — it writes small programs against your tools instead of calling them one at a time.
Secrets live in a per-user vault and never appear in the conversation.
Works across channels — the same connected servers serve web chat and inbound messages.
04
Native web search, code interpreter & image generation
toolsmobile
Web search migrated from a bolted-on tool to the provider's native one, and code interpreter plus image generation joined it. Charts and generated images render in web chat and are delivered through to the iPhone app.
05
Agent packs became buildable artifacts
packsnix
Professions ship as installable packs — soul, skills, and automation templates in one portable directory format. A serializer makes packs fully round-trippable, and a Nix realization layer turns every pack into a buildable flake, pinned toolchain included. The first production pack was reverse-engineered from the agent that runs this company: 50 skills and 11 automations, pushed back through the standard pipeline.
Core pack plus five profession packs authored against the new format.
Automations can ship real scripts (scriptPath) and pull custom packages via flake refs.
Import validation hardened across three review waves: slugs, versions, nix refs, manifest content, and cron edge cases.
06
Design-system groundwork
designin review
A forensic study of a reference-grade product UI — twelve reports and a construction bible — plus a dark-first token layer and full restyles of the landing, docs, chat, and UI primitives built on it. This one is still in review, which is why the page you're reading looks the way it does.
07
The mobile app grew up
iosandroidexpo
The iPhone app learned attachments end to end — attach from the library, paste an image, or share into the composer from any app — and Android arrived with build and Play-store delivery parity. The share sheet now asks whether you want a new thread or an existing one.
Composer polish: one-tap reasoning menu, send button on the input line, paste folded into the “+” menu.
ChatGPT quota usage is visible in the app, and the TestFlight pipeline was unblocked (Xcode 26 for SDK 56).
08
Chat & agent UX
webux
A reasoning-effort selector next to the model picker, the sidebar split by channel, ChatGPT quota usage surfaced in web and mobile, and a redesigned reasoning and tool-call display. Inference errors are now classified — when your quota runs out, the agent says so on every channel instead of failing silently.
git log --since="07-07 10:00"127 commits
01:4769133e8feat(mobile): iOS share-sheet target picker (New thread or existing)
01:31056bc3bfeat(agent): classify inference errors + surface quota-exhausted to all channels
00:42522d5dbfeat(tools): native code_interpreter + image_generation, mobile viz, image delivery
18:000744e68refactor(chat): drop CHAT_RUNTIME flag
14:035259672feat(avatars): nix realization layer — every pack is a buildable flake
10:58dedecfcfeat(mobile): attach, paste, and share files into the chat composer
… 115 more
Shipped means pushed and reviewed — a few of these are still rolling out behind flags or making their way through final review.
Groundwork weekend.
The quieter weekend that made the sprint possible: a walking skeleton of the new chat runtime, and a landing page that finally tells the truth about the product.
A Convex httpAction learned to run a full chat turn — streaming spike, resumable-stream store, then tools, memory, the bash bridge, Google Drive mounts, and file uploads, one commit at a time.
The landing page was repositioned around what nicoolAI actually is — a virtual assistant — and rebuilt around the brand green with real product content instead of a generic task list.
The landing design became the single source of truth for theme tokens, so the dashboard and docs draw from the same palette.
Android support began its climb to iOS parity: builds, Play delivery, and shared composer code.
All of this works for you today.
Sign up in a minute and hand nicoolAI the first thing on your plate.
