Builder · Level 2 of 3
This is a desk-research first look, drawn from Scarlett’s own site, its FAQ, and outside reviews — not a hands-on trial, since the product only became publicly available in July 2026.
Scarlett is a new AI agent that works inside Slack, and in a more limited way inside iMessage, as a coworker rather than a chat window. Instead of answering a question and leaving the rest to you, it logs into a cloud computer of its own, connects to the tools your team already uses, and finishes the task: updating a CRM record, pulling last month’s numbers into a report, or shipping a PDF a teammate asked for. It launched publicly on Product Hunt on July 10, 2026, so here’s what it actually does, what it costs, and where the access tradeoffs sit before you install it in a real workspace.
What is Scarlett, exactly?
Scarlett is built by Cracked AI and installs as a Slack app. The company’s pitch is specific: it’s a coworker, not a chatbot. It keeps a persistent cloud workspace where it can write and run code, and it uses that workspace to finish tasks rather than just describe how you’d do them yourself.
Under the hood, Scarlett runs on Claude and routes different parts of a task to different models — Opus for conversation, Fable for planning, and Claude 5.6 for coding, picking whichever fits the job instead of running everything through one model. That multi-model routing is a common pattern among the newer wave of “AI coworker” products.
The basic interaction is simple: mention @Scarlett in a Slack channel with a request, and it goes and does the work using the tools you’ve already connected, then reports back with a finished output instead of a list of instructions for you to carry out.
Where does Scarlett live, and what do you need before you start?
Scarlett is Slack-first. It works as an app inside your existing Slack workspace, sitting in channels and threads the way a coworker would. A more limited iMessage version also exists for one-on-one use outside Slack, though outside reviewers note it carries tighter role-permission limits than the Slack version, since standard messaging protocols weren’t built for team-level access control.
Before installing it, you need:
- A Slack workspace, with permission to install apps (usually a workspace admin has to approve this)
- The accounts you want Scarlett to touch already set up — Stripe, Notion, HubSpot, Google Drive, or whichever tools apply
- A decision about which channels Scarlett can see, since it only reads channels it’s explicitly invited to
Setup itself is short:
- Install Scarlett from the Slack App Directory, which authorizes it for your workspace.
- Connect the specific tools you want it to use through its Connections page. Credentials go through an authenticated connections screen, not through chat.
- Invite Scarlett into the channels where you want it to work.
- Mention @Scarlett with a task the way you’d ask a colleague — “pull our ad spend data and compare it to last month” or “build me a revenue dashboard.”
The company puts setup at two to three minutes once you have admin access and know which tools you’re connecting first.
What can Scarlett actually do?
The examples on Scarlett’s own site and FAQ line up with the CRM-and-reports framing in its launch pitch:
- Update CRM records (Salesforce, HubSpot) after a call or meeting
- Pull numbers from Stripe, QuickBooks, or a spreadsheet into a written report
- Draft a marketing strategy document or a PDF from a meeting summary
- Turn a week of Slack, Asana, and Google Drive activity into a status digest
- Watch a Google Drive folder for new contracts and flag anything worth a second look
- Sweep Gmail for threads without replies and draft the follow-up
It also ships a library of pre-built recurring workflows you can turn on directly: a Monday-morning business report pulling from Stripe and HubSpot, a daily support-ticket triage from Zendesk and Intercom, a weekly customer-feedback theme report pulled from support and email. These run on a schedule and post the result into Slack without you asking each time.
Scarlett claims connections to more than 3,000 tools, including Salesforce, HubSpot, Linear, Notion, Jira, Stripe, GitHub, Google Drive, Gmail, and Zendesk. If a tool you use isn’t already supported, the vendor says it will build a custom integration; confirm that directly with the company before you commit if it’s the deciding factor for you.
What does Scarlett cost?
The published plan is straightforward. The Team plan is $50 a month and includes 25,000 credits a month, all 3,000+ integrations, scheduled and recurring tasks, and higher concurrency limits than a trial account. Credits are consumed by the work Scarlett does, so a workspace running several recurring reports will burn through the monthly allotment faster than one asking occasional one-off questions. Enterprise pricing is custom and adds single sign-on, role-based access, invoicing, a security review, and dedicated onboarding.
What’s verified here: the pricing and plan details above come directly from Scarlett’s own site as of late July 2026. Credit-based pricing for AI agents tends to shift as usage patterns settle, so treat the exact numbers as current at the time of writing, and check the pricing page again before committing budget.
What happens to your data when an AI agent gets this much access?
This is the part worth slowing down for, since Scarlett by design needs broad access to work: your CRM, your inbox, your calendar, your file storage, sometimes your billing tools.
A few specifics from the vendor’s own security documentation:
- Each workspace gets isolated cloud infrastructure, so one customer’s Scarlett instance isn’t sharing compute with another’s.
- Data is encrypted both in transit and at rest, and the company states it does not train its models on customer data.
- OAuth tokens are held by a separate connection proxy rather than passed to the AI model directly; custom credentials are encrypted with AES-256-GCM and only decrypted server-side at the moment an authorized action runs.
- Scarlett only reads Slack channels it has been explicitly invited to, and it doesn’t have blanket access to a workspace’s full message history.
- Sensitive actions, like sending an email or deploying a change, wait for your confirmation instead of firing automatically.
- You can request a full data wipe at any time through account settings or support.
None of that removes the underlying tradeoff: you’re handing a third-party AI agent standing access to systems that, in the wrong hands, could touch customer data, financial records, or your team’s private conversations. Scarlett’s own product copy is upfront that it “may generate inaccurate responses” and asks you to review important outputs before acting on them. Treat that as a real instruction, not boilerplate: a CRM update or a sent email based on a misread request is a normal failure mode for this category of tool, not a hypothetical one.
If your business handles regulated data, such as health records, payment card data, or anything under a strict data-processing agreement with a client, get your own security and compliance review before connecting those systems, regardless of what any vendor’s marketing page claims.
How does Scarlett compare to other AI coworkers?
Scarlett’s own site publishes a benchmark comparing itself to two named rivals on an internal quality score and average response time:
| Product | Quality score (vendor benchmark) | Avg. response time |
|---|---|---|
| Scarlett | 92.1 | 9.8s |
| Claude Tag | 90.2 | 25s |
| Viktor | 89.2 | 14s |
That’s a vendor-run comparison published on Scarlett’s own site, not an independent benchmark, so treat the specific numbers as a marketing claim rather than settled fact. It’s still useful context for what Scarlett optimizes for: fast, high-quality task completion, not conversational depth.
The more useful distinction is structural. Tools like ChatGPT and Gemini are still mostly conversation-first: you ask, they answer, and turning that answer into a finished action in another tool is on you. AML’s guides to organizing work with ChatGPT Projects and everyday tasks with Google Gemini show what that looks like in practice. Scarlett and similar “AI coworker” products close that last gap: the model plans the task and then executes it directly in your connected tools, without you copying an answer somewhere else. That’s a meaningfully different category, even when the underlying language model doing the reasoning is a similar generation of technology.
Should you try Scarlett yet?
Try it if:
- You already run a Slack-first workflow and want one of the recurring report or triage workflows above, close to as-is
- You’re comfortable connecting business tools to a third-party agent and reviewing its outputs before they go live
- You want to test an AI agent that executes tasks end to end rather than just drafting text for you to finish
Hold off if:
- Your team can’t tolerate an occasional wrong output on customer-facing or financial tasks without a review step
- You need on-premises or fully self-hosted control over where your data lives
- You’re not ready to spend the first week narrowing scope (which channels, which tools, which actions need your sign-off) before trusting it with recurring work
Verdict
Scarlett is a genuinely new shape of Slack tool: less a chatbot, more a teammate with its own login to your CRM and calendar. The pitch is credible and the setup is short, but it’s a July 2026 launch with a vendor-run benchmark and no independent long-term track record yet. Start it on a low-stakes recurring task, like a Monday status digest or a support-ticket triage, before handing it anything that touches money or customer data directly, and keep the confirmation step on for anything that sends, deploys, or bills.