Loading...
Anthropic's computer use requires a Pro or Max plan and is unavailable on Team and Enterprise at any price. Here is the free route on macOS using an MCP server, including the daily limits and the failure modes you will hit.
Give Claude, Codex, or any AI assistant the ability to see your screen and automate any task. Runs locally on your Mac or PC.
You can get Claude computer use free on macOS without a Pro or Max plan by running an MCP server that drives the screen and pointing your Claude client at it. OpenOwl's free tier covers 50 tool calls per day and asks for no credit card.
Skipping Pro or Max costs you Anthropic's own implementation. You keep the ability to let Claude see and control your Mac through a different path. Below is that path, its limits, and the places it breaks.
Anthropic states the conditions in its own documentation:
Computer use is a research preview on macOS that requires a Pro or Max plan. It is not available on Team or Enterprise plans. It requires an interactive session, so it is not available in non-interactive mode with the
-pflag.
Those two sentences carry three separate limits.
First, the plan gate. Anthropic switches computer use on for Pro and Max subscribers, and someone on a free Claude account never gets the option.
Second, the plan exclusion, which catches more people than the gate does. Anthropic blocks computer use on Team and Enterprise plans at any price, so paying more moves you further from the feature.
Third, the interactive-session requirement. The -p flag gives you nothing here, so you cannot drive screen control from a cron job or a CI step.
Anthropic documents one more constraint outside that quote. Computer use takes a machine-wide lock and permits one session at a time. As of v2.1.195, finishing a task leaves that lock held. You free it by exiting the session, so a Claude window you forgot about will block your next run.
People conflate two Anthropic products here. Claude Code's CLI computer use runs on macOS, and Anthropic rules out Linux and Windows for it. The Claude Desktop app covers macOS and Windows.
You can see the mechanism in how computer use authenticates. Anthropic requires claude.ai authentication and ships no support for Amazon Bedrock, Google Cloud's Agent Platform, or Microsoft Foundry.
Large companies buy Claude through those three providers. A company routes its AI spend through an existing cloud contract for billing, data residency, and audit reasons. Anthropic never wired computer use into that path, so the exclusion follows from the architecture.
An engineer sitting on a company Enterprise seat has no upgrade path. Two options remain. One is a personal Pro or Max subscription, which means pushing work data through a personal claude.ai account, and many security teams reject that. The other is running the screen-control layer outside Anthropic's plan structure.
OpenOwl runs as an MCP server. That architecture is what gets you around the plan gate.
MCP (Model Context Protocol) is an open standard for connecting AI assistants to external tools. Anthropic ships computer use as a feature inside a product, so Anthropic decides which plans see it. An MCP server is a separate process on your own machine that advertises tools over stdio, and any client that speaks MCP can call them without a vendor checking your plan first.
The path looks like this:
Claude Code / Codex CLI / any MCP client
|
| stdio transport
v
owl (process on your Mac)
|
v
macOS Accessibility + Screen Recording APIs
|
v
your screen
Two consequences matter. Your screenshots stay on your machine, because owl reads the screen through macOS APIs and hands the results back to the client you already trust. You also avoid the machine-wide lock, since each MCP client spawns its own owl subprocess.
You keep your choice of assistant too. Claude Code and Codex CLI both have tested configurations, and any other MCP client uses the generic stdio config. Switch models a year from now and your automation layer survives the switch. For how this approach compares with the other tools in the category, see our breakdown of the best computer use agents.
The free tier meters tool calls at 50 per day and gives you the full tool set. OpenOwl ships 40 tools across 14 modules: screenshots, input, accessibility-tree queries, OCR, window management, visual diffing, and workflow record and replay.
A few of those tools do more work than their names suggest. screenshot runs OCR on every capture and appends on-screen text coordinates (capped at 50 phrases) plus an accessibility element summary (capped at 30). click captures a 400x400 region before and after, reports whether the pixels changed, and returns the focused element so the model can confirm the click landed. click_element runs a three-tier cascade: accessibility tree first, then fuzzy match, then OCR.
The two approaches compare like this:
| Constraint | Anthropic computer use | OpenOwl free tier |
|---|---|---|
| Plan requirement | Pro or Max | None, no credit card |
| Team / Enterprise plans | Not available | Works |
| Cloud provider routing | claude.ai auth only | Not applicable |
Non-interactive (-p) use | Not available | Works |
| Concurrent sessions | One, machine-wide lock | One process per client |
| Assistant choice | Claude only | Any MCP client |
| Daily limit | Plan usage limits | 50 tool calls, 10 per minute |
| Platform | macOS (CLI), macOS and Windows (Desktop app) | macOS on Apple Silicon and Intel |
Read the platform row twice, because getting it wrong wastes an afternoon. OpenOwl ships for macOS. Apple Silicon and Intel Macs both work, and no Windows binary exists. If you need Windows screen control, use the Claude Desktop app.
The installer takes your API key, drops the binary in place, and writes the key file:
curl -fsSL https://openowl.dev/install.sh | bash -s -- owl-xxxx-xxxx-xxxx
Generate a key first on the OpenOwl quick-setup page. It creates one against a free account and prints these commands with your key filled in.
If you prefer not to pipe a script into bash, the npm path does the same work in four steps:
npm install -g openowl
mkdir -p ~/.openowl
echo "owl-xxxx-xxxx-xxxx" > ~/.openowl/api.key
claude mcp add owl --transport stdio -s user -- owl
Codex CLI users add this to ~/.codex/config.toml in place of the claude mcp add line:
[mcp_servers.owl]
command = "owl"
Two permissions remain, and macOS enforces both. It blocks any process from reading the screen or driving other apps until you grant them in System Settings > Privacy & Security:
Grant both, then restart the server. macOS does not re-evaluate permission grants for a process that is already running, so you have to restart it. The server runs a preflight at startup, which puts both dialogs in front of you when you launch it.
If screenshots come back black, suspect a missing or stale Screen Recording grant. Restart the server before you debug anything else.
Fifty calls per day sounds larger than it is. Each tool call counts one action, so a minimal act-and-verify loop spends three: capture the state, click the target, capture again to confirm. Fill a six-field form with a check between fields and you have spent a real share of the day's budget in one run.
The free tier also caps you at 10 calls per minute. Both counters reset at midnight UTC. Validation is strict, and a missing or invalid API key blocks the server outright.
Cost per call is uneven. A single click includes a forced 0.5s sleep plus two 400x400 region captures plus an accessibility query, which makes it one of the more expensive calls in the set.
You hit four limitations regardless of tier, and knowing them before you build saves rework:
find_element and list_elements, so you fall back to OCR for anything inside a page.window_title or capture a region.--force-renderer-accessibility.jabswitch -enable or fall back to OCR.Build one habit into your first workflow. Never estimate coordinates by eye from a screenshot: captures downscale to 1280px wide, so your guess lands in the wrong place. Use the coordinates the tools return.
The OpenOwl pricing page lists the current tiers, including Pro at $19.99 per month.
Anthropic's built-in computer use is the better choice in several situations, and pretending otherwise would waste your time.
Pay for Max when you want one vendor answering for the whole path from model to click. An MCP server plus a client is two projects, and integration problems land in the seam between them, where you pay for them with your own time. Pay for Max as well when Windows is your target, since the Claude Desktop app covers Windows and OpenOwl ships no Windows binary.
If you hold the subscription already, computer use comes included, and a second tool for the same job buys you portability across assistants and little else.
Take the MCP route when your Team or Enterprise plan puts Anthropic's feature out of reach at any price, when you need screen control from a non-interactive script, when you want the automation layer to outlive your current model choice, or when you want to try the idea before you commit to a line item.
Both can run on the same machine. OpenOwl on a free account costs you nothing, so run it for a week and you will know whether screen control earns a subscription.
Anthropic's built-in computer use requires a Pro or Max plan, and Anthropic blocks it on Team and Enterprise plans. You can get screen control free on macOS by running an MCP server such as OpenOwl, whose free tier allows 50 tool calls per day with no credit card. Your Claude client's own access terms are separate and still apply.
Yes. OpenOwl runs as an MCP server on your own Mac, so Anthropic's plan restrictions never apply to it. That matters for Team and Enterprise seats, because Anthropic makes computer use unavailable on those plans at any price and offers no path through Bedrock, Google Cloud's Agent Platform, or Microsoft Foundry.
No. OpenOwl ships for macOS, on both Apple Silicon and Intel. Anthropic's Claude Desktop app covers Windows for its own computer use, so use that when Windows is a hard requirement.
The server stops serving tool calls until the counter resets at midnight UTC. A free key also carries a rate limit of 10 calls per minute, and a fast loop reaches that ceiling first. Upgrading raises both.
Anthropic's version blocks that. Its documentation states that computer use requires an interactive session and does not work in non-interactive mode with the -p flag. An MCP server carries no such restriction, so a client running headless can still call the tools.
Generate a free API key on the quick-setup page, run the install command, grant the two macOS permissions, and restart the server. Then ask Claude to take a screenshot of your desktop. If the capture comes back, both grants landed, and you have spent nothing to find out whether this solves your problem.