Claude Cowork Limitations in 2026: What It Still Can't Do
Claude Cowork is genuinely capable. It reads and writes files in folders you grant, browses the web and fills forms in Chrome, builds spreadsheets and slide decks, runs multi-step plans that take an hour or more, and — since the summer 2026 update — runs remotely on Anthropic’s servers so scheduled tasks fire with no device online. The old “it only works when your laptop is awake” complaint is out of date.
But if you’re evaluating Cowork to run real, recurring work, the walls show up in specific places. Here’s an honest map of them — verified against Anthropic’s own documentation — and where a different architecture solves each one.
It’s metered, and multi-step work burns through the meter fast
Cowork isn’t a flat-rate tool. It’s bundled into your Claude plan (Pro around $20/mo, Max 5x around $100/mo, Max 20x around $200/mo; the free plan doesn’t include Cowork at all), and every session draws down your usage allocation. Model responses, tool and skill calls, image generation, and browser tasks all count toward consumption.
The catch is that Cowork’s whole value proposition — long, multi-step, autonomous work — is exactly what eats the meter. As Anthropic’s guidance and third-party testing note, Cowork tasks consume far more of your allocation than standard chat because multi-step work is token-intensive, and there are quotas on action steps and timeouts on long-running tasks. Run a few hour-long browser jobs and you can hit your window’s ceiling. This is a fine model for interactive, occasional work. It’s a harder fit for “do this every day forever,” where the cost scales with how much the agent actually works.
No inbox of its own — you always start the conversation
This is the structural one. Cowork is a general agent you prompt. You open Claude, describe an outcome, and it goes. Even scheduled tasks are prompts you wrote earlier, running on a cadence you set.
What it isn’t is a teammate other people can reach. There’s no email address a client, a lead, or a colleague can write to and get a response from the agent. Nobody outside your Claude account can hand it work. If the trigger for a task is “someone emailed me,” “a lead filled out the form,” or “a calendar invite landed,” Cowork has no native way to know that happened — you have to notice it and then go prompt Cowork about it.
That gap is the entire premise behind Carly, which gives each agent its own name, email address, and memory. People email or text Carly directly and it replies and does the work underneath — the reachability that a prompt-driven desktop agent structurally can’t offer.
Sending email means Microsoft 365 — and only Microsoft 365
Cowork can send email, but there’s a real constraint on how. The path that actually sends is the Microsoft 365 connector. If you live in Google Workspace, that route is closed for sending. And even on M365, the write tools carry restrictions worth knowing before you rely on them:
- No attachments. Sending, forwarding, and drafting all reject any message that has an attachment.
- Per-user caps on writes, sends, and recipients.
- No deleting files or folders in OneDrive or SharePoint, and no access to files stored locally on your device.
None of this makes Cowork useless — it makes it a supervised assistant for a specific stack, not a general-purpose sending-and-filing agent across whatever tools you actually use.
You’re still the supervisor
Cowork gives you three approval modes: manually approve each step, automatically approve (with Claude self-blocking anything it judges unsafe), or skip all approvals. Anthropic is refreshingly blunt that no mode replaces your judgment — web content is a primary vector for prompt-injection attacks, some sites require approval on every single action, and “skip all approvals” removes a safety layer rather than adding convenience.
The honest read: Cowork is designed to be watched. When you scope tasks into discrete, auditable steps and keep an eye on it, it’s reliable. When you point it at open-ended, poorly defined, or long-running work and walk away, the likelihood and impact of errors both rise. That’s the right posture for a powerful tool — but it means Cowork is closer to a copilot you drive than a colleague you delegate to and forget.
Connectors cover the mainstream; the long tail needs setup
Cowork’s integrations are GUI-configured connectors — Slack, Linear, Google Calendar, Gmail, GitHub, and the other mainstream names — plus MCP support that’s limited to those connector-style integrations. That’s solid coverage for common tools. But anything off that list means standing up your own MCP server or extension, and there’s no toggle to paste an API key and reach an arbitrary app.
For a lot of solo operators and small teams, the tools that matter are exactly the long tail: a niche CRM, a legal or property-management platform, an accounting app, a scheduling system. If it’s not a first-class Cowork connector, you’re doing integration engineering.
Browser automation is impressive but session-bound
Cowork’s browser control — clicking, typing, filling forms in Chrome — is one of its most demo-worthy features. It’s also where reliability is hardest, because it’s driving live web UIs. Beyond prompt-injection exposure, the sandbox restricts Cowork’s ability to persist state across sessions, and long browser jobs run into timeouts and action-step quotas. Multi-step tasks that touch several sites are precisely the ones most likely to need a human to step in.
Where each gap gets closed
Most of these limits trace back to one design choice: Cowork is a general, prompt-driven agent metered by your Claude usage. If what you actually want is an agent other people can reach, that fires on events instead of prompts, and that reaches your whole tool stack, that’s a different shape of product.
That’s the shape Carly takes. It runs in the cloud on triggers — inbound email, a calendar invite, a Slack message, a form submission, a schedule — rather than waiting to be prompted. It gives each agent its own inbox people can write to. It reaches 260+ native connectors across 45+ categories, and for anything else with a public API you paste your own key at dashboard.carlyassistant.com/integrations — so the long tail isn’t a wall. And on pricing, the Zapier-style Workflows are free and unlimited (non-AI steps aren’t metered), with AI agents from $35/month, so building more automation doesn’t automatically mean a bigger bill.
Cowork is a strong interactive coworker for the person sitting in front of it. The gaps above are what to weigh if you need an agent that works when nobody’s watching and that other people can reach directly. If that’s the job, it’s worth comparing against dedicated alternatives built for it.
Ready to automate your busywork?
Carly schedules, researches, and briefs you—so you can focus on what matters.
See what people say
"Before Carly, I relied on a Calendly link, but the whole process felt impersonal and not very professional. Carly changed that by handling all the back-and-forth, so I'm no longer stuck in endless email threads trying to line up schedules.
Now Carly reaches out to candidates, shares my real-time availability, lets them pick a slot, then sends a Zoom link and drops it straight into my calendar. She sends reminders to both of us before each call, which has significantly reduced no-shows and last-minute confusion.
On top of scheduling, Carly acts like a full executive assistant, sending me my schedule the night before so I can prepare for each call. It reminds me of the old x.ai assistant, but Carly is noticeably smarter, faster, and better suited to my healthcare recruitment business."
