Caspar David Friedrich, Two Men Contemplating the Moon, ca. 1825–30. The Metropolitan Museum of Art, public domain.
Between January and March 2026, Anthropic shipped four releases that change what its tools can do for a small team: Claude Cowork for non-developers in January, Claude Opus 4.6 with a one-million-token context window in February, and Auto Mode for Claude Code and Computer Use in Cowork in March. Together they move Claude from a chat window to an agent that edits files, runs code and operates a desktop.
I use Claude Code every day across four AI labs in Jamaica, and I have tested the other major AI coding tools. For most Caribbean teams, the right first step is Claude Code for developers and Cowork for one repetitive office task, with a clear rule on what data the agent may see.
Claude Code went from terminal tool to platform in 13 months
Claude Code launched in February 2025 as a command-line tool. It reads a codebase, writes code, runs tests, fixes bugs and handles git from the terminal.
- February 2025: initial release as a terminal coding tool
- May 2025: general availability alongside Claude 4
- August 2025: Claude for Chrome extension for browser control
- October 2025: web version and iOS app
- January 2026: annualised revenue reported at US$1 billion
- March 2026: Auto Mode launches; annualised revenue reported above US$2.5 billion
Auto Mode approves routine actions and stops destructive ones
Before Auto Mode, Claude Code asked permission for every file write and terminal command, or you switched approvals off entirely. Auto Mode uses a safety classifier to sit between those two:
- Routine actions run without a prompt: file reads, code edits, test runs, linting and ordinary git commits.
- Destructive actions still need a human: force pushes, database drops,
rm -rfand similar commands. - Context changes the decision: deleting a test fixture and deleting a production database are treated differently.
In my labs it removed most of the clicking that made long tasks tedious, while still stopping on the commands I would want to see.
Channels: Claude Code from Discord and Telegram
Channels lets a team give Claude Code tasks through Discord or Telegram. It is Anthropic's managed alternative to open-source agents such as OpenClaw. A developer in Kingston can assign a task in a channel and a designer in Montego Bay can review the result in the same thread.
Cowork brings the same agent to non-developers
Claude Cowork, launched in January 2026 as a research preview in the Claude Desktop app, gives non-technical users a graphical way to work with Claude on files. It differs from ordinary chat in three ways:
- Folder access. You choose a folder on your computer that Claude can read and change, so it works on your actual files.
- File work. It creates, edits, renames and organises files in that folder: a report, a cleaned spreadsheet, a sorted project directory.
- Connectors. MCP connectors for Google Drive, Google Calendar, Gmail, DocuSign, Apollo, Clay, Outreach, SimilarWeb, MSCI, LegalZoom, FactSet, WordPress and Harvey let Claude use those services through their APIs.
Connectors matter because an API call is faster and more reliable than an AI reading pixels on a screen, and it leaves a cleaner log. Cowork uses a connector when one exists and falls back to screen control only when it does not. Boris Cherny, who created Claude Code, has said Cowork was "mostly built by Claude Code."
Where Cowork fits in a Caribbean office
- An accounting firm in Kingston: sort client tax documents, draft client emails and update working spreadsheets from Google Drive.
- A marketing agency in Bridgetown: maintain a WordPress content calendar and draft social posts against Google Calendar deadlines.
- A law firm in Port of Spain: summarise case files and track contracts sent for signature through DocuSign.
- A sales team anywhere in CARICOM: update prospect lists and sequences through the Apollo and Outreach connectors.
Computer Use lets Claude operate a Mac
Since March 2026, Claude can control a macOS desktop through Cowork: opening files and applications, pointing, clicking and typing. Key facts:
- Available on Pro (US$20 a month) and Max (US$100 or US$200 a month) plans.
- macOS only during the research preview.
- Works through Cowork in the Claude Desktop app, and through Dispatch, which lets you send a task from the Claude mobile app to Cowork on your computer.
- Anthropic's models now score 72.5 percent on OSWorld, a benchmark for agents using real computer environments, up from under 15 percent for early computer-use models.
With Dispatch you can type on your phone, "open the Q4 sales report on my laptop, update the March numbers and email it to the team," and Cowork carries it out on the desktop. The computer has to be on, awake and running the Claude app.
The model underneath: Claude Opus 4.6
Claude Opus 4.6, released on 5 February 2026, has a one-million-token context window, roughly 750,000 words of English. That lets Claude Code hold a whole mid-sized codebase in view, and lets Cowork work through long documents and large folders without losing track. Sonnet 4.6 followed on 17 February as a faster, cheaper option for tasks that do not need Opus.
Where your data goes
Cowork and Claude Code run on your computer, but the model does not. Every file, screenshot or web page the agent reads is sent to Anthropic's servers for processing, and the answer comes back. Do not assume anything stays on your machine. Point Cowork at a working folder that holds only what the task needs, keep client personal data out of it unless your contract and your country's data protection law allow it, and use a business plan where one is available.
The market reaction and Caribbean BPO
IBM shares had their worst day since October 2000 after Anthropic published a post on using Claude Code to modernise COBOL, and enterprise software stocks fell after the Cowork announcement. Investors are pricing in agents doing work that software services firms bill for.
Jamaica's BPO sector employs tens of thousands of people. Agents equipped with these tools can offer higher-value services, and clients using the tools directly may need fewer outsourced staff. Firms that retrain now for supervision, quality checking and integration work will keep more of their contracts.
Where these tools still fall short
Computer Use is slow compared with a connector and still misclicks, so it suits tasks where a mistake is easy to spot and undo. Auto Mode's classifier is new, and I do not yet know how it behaves on unusual setups such as a shared server with production credentials in the environment. I keep production keys off any machine where Claude Code runs unattended, and I would advise the same until there is a longer track record.
What Caribbean developers and businesses should do next
- Developers: install Claude Code on a Pro plan and add a
CLAUDE.mdfile to your main repository describing your conventions, build commands and test commands. Claude Code reads it at the start of every session. - Developers: write or update the tests for the next feature before asking Claude Code to implement it. Clear tests give the agent a definition of done.
- Office managers and team leads: pick one weekly file task, such as a report built from the same spreadsheets, and run it in Cowork on a dedicated folder for four weeks. Compare the time taken with your usual process.
- Business owners: before connecting Gmail, Drive or a CRM, write down which data may be sent to an AI provider and check it against your country's data protection law.
- Mac users trying Computer Use: start with tasks you can check in a minute, such as renaming files or filling a form you review before submitting, and never let it act on banking or payment screens.
Frequently Asked Questions
Does Anthropic train its models on what I do in Claude Code or Cowork?
It depends on your plan and settings. On commercial plans such as Team, Enterprise and the API, Anthropic states that it does not train on your content by default. On consumer Free, Pro and Max plans, users choose in their privacy settings whether their chats and coding sessions may be used for training. Check the setting under your account's privacy options before using client material.
Is Claude available in every Caribbean country?
Claude is available in most Caribbean countries, but not all; Anthropic publishes a list of supported countries and regions on its website. Payment can also be a hurdle, since subscriptions need a card that works for international online payments. If your local debit card is declined, a credit card or a bank card enabled for online foreign transactions usually works.
What is the difference between a Pro plan and paying through the API?
A Pro or Max subscription charges a flat monthly fee with usage limits that reset over time, which suits one person using Claude Code or Cowork daily. The API charges per token used, which suits automated jobs and teams building products, and costs can rise quickly on large codebases. If you use the API, set a monthly spend limit in the Anthropic Console before you start.
What is the most common mistake new Claude Code users make?
Giving it a large, vague task on an unfamiliar codebase with no tests and no CLAUDE.md file. The agent then guesses at conventions and produces changes that are hard to review. Start with small, well-defined tasks, ask it to explain its plan before editing, and review each diff before committing.
Can I use Claude Code on Windows?
Yes. Claude Code runs on macOS, Linux and Windows, including through Windows Subsystem for Linux. Cowork runs in the Claude Desktop app, and Computer Use inside Cowork is macOS only during the research preview. Windows users can still use Cowork's folder and connector features.
"Between January and March 2026, Anthropic gave developers Auto Mode, gave office workers Cowork, and let Claude operate a desktop. A developer in Kingston can use all of it on the same day as a developer in San Francisco. Access is no longer the gap. Adoption is." Adrian Dunkley, AI Boss