Which Mac You Actually Need to Run Claude Code

The official requirements for Claude Code, as the setup guide on this site notes, are macOS 13 or later and 4 GB of RAM. Both numbers are true and neither one helps you buy a laptop. They describe what it takes to launch the CLI, not what it takes to work in it for six hours with three sessions running, a dev server up, and a browser holding forty tabs.

The useful version of the question is narrower. Claude Code does almost nothing computationally on your machine. The model runs in Anthropic's data centre. What runs on your Mac is a Node process, a lot of file reads, and whatever the agent decides to spawn: ripgrep across the repo, a test suite, a build, a container. Your Mac is not thinking. It is doing I/O and running your toolchain, faster than you would have, and more often.

That changes which specs matter.

What actually consumes your Mac during an agent session

Open Activity Monitor during a long Claude Code run and the process itself is unremarkable. A single session sits in the hundreds of megabytes. It is the tail that costs you:

• The editor, which is usually Electron and usually the largest single process on the machine, along with whatever else you have open to read what the agent produced
• Your dev server, watcher, and whatever it reloads on every file the agent touches
• The test suite the agent runs on its own initiative, sometimes several times in a row
• Docker or Colima, if your project needs it, which reserves memory whether or not you are using it
• The browser, which is not part of your toolchain but is always there anyway
• Every additional Claude Code session, each with its own context, its own spawned processes, and its own reason to run the tests

Run one session and none of this is a problem on any modern Mac. Run three across separate worktrees, which is the entire point of using an agent and the habit that separates it from an in-editor assistant, and you are running three toolchains at once. That is the real workload, and it is a memory problem, not a compute problem.

How much RAM you actually need

RAM	What it handles	Honest verdict
8 GB	One session, one editor, a small project, few tabs	Works, with swap. Painful with containers
16 GB	Two or three sessions, editor, dev server, browser	The sensible floor for daily agent work
24 GB	Several sessions plus Docker, or a large monorepo	Comfortable. Where most people stop needing to think about it
32 GB+	Heavy containers, multiple large repos, local models alongside	Buy it if you already know why

Every current MacBook Air ships with 16 GB of unified memory as standard, so a new machine clears the bar without a configuration decision. The question only arises on older hardware and on the used market, where 8 GB machines are common and cheap for a reason.

Apple Silicon handles memory pressure gracefully. Swap on an SSD is fast enough that a machine under pressure feels sluggish rather than broken. Two things follow from that. First, an 8 GB Mac will run Claude Code perfectly well for a single session on a modest project, and anyone telling you it is unusable is overstating it. Second, sustained swapping writes constantly to storage you cannot replace, because it is soldered to the board. Living permanently in swap is a slow tax on a machine you plan to keep for five years.

RAM is also the one spec you cannot change after purchase. Storage you can work around with an external drive. Memory is welded to the chip. Buy it once.

Does the chip generation matter?

Less than you would expect. Between an M1 and an M5 running the same agent session, the difference shows up in your builds and your test suite, not in the agent. Compile times shrink, the tests finish sooner, and the loop tightens. That is worth real money if you are waiting on builds all day, and close to nothing if you are not.

Memory bandwidth, the spec that separates the Pro and Max chips in every benchmark article, is almost irrelevant here. Bandwidth determines how fast a model generates tokens, and your model is not on this machine. It matters enormously if you plan to run models locally instead, through Ollama or LM Studio, which is a genuinely different spec sheet and one covered separately in RefurbMe's guide to picking a Mac for local AI. For a cloud agent, ignore it.

Air or Pro: the fan is the real difference

The MacBook Air has no fan. Under a short burst it matches a Pro. Under sustained load it heats up, throttles, and settles into a slower steady state.

A long agent run is sustained load. Not the agent, but what it triggers: repeated builds, full test suites, a watcher rebuilding on every write, all of it in sequence for twenty minutes at a time. On an Air, the second half of that run is slower than the first. On a Pro, it is not, and the machine stays quiet enough that you stop noticing.

This is the one place where the extra money buys something you feel every day rather than something you see in a benchmark. If your project builds in four seconds, get the Air and spend the difference on RAM. If you compile anything substantial, or you leave agents running while you do something else, the fan earns its keep.

Is an M1 still fine in 2026?

Yes, with one condition: 16 GB.

An M1 MacBook Air with 16 GB is a completely reasonable Claude Code machine in 2026. It runs current macOS, it handles two sessions and an editor without complaint, and it is silent. It is also six years old, which puts it in the part of the used market where prices have collapsed, and it caps at 16 GB, which means it will never be more than a two-or-three-session machine.

The 8 GB M1, which is the configuration most people bought and therefore the one flooding the secondhand listings, is the one to avoid. It is the cheapest Apple Silicon Mac you can buy and it is cheap because everyone who tried to do modern work on it has already replaced it.

What to buy, by budget

Budget	Machine	Why
Tightest	M1 or M2 Air, 16 GB	Silent, capable, and the best price per useful year on the market
Middle	M3 or M4 Air, 16 GB	Faster builds, better display, still fanless
Comfortable	M3 or M4 Pro 14-inch, 18 to 24 GB	The fan, the screen, and headroom for containers
No constraint	M4 Max or M5 Max, 36 GB+	Only if you are also running local models

Note what is absent from that table: the newest chip. A current M5 Air starts at $1,299. The M3 and M4 machines that were flagships eighteen months ago do the same job for an agent workflow and sell for a fraction of that, because Apple's pricing tracks release date far more closely than it tracks capability.

This is where buying refurbished stops being a compromise and starts being the obvious move. Apple's own Certified Refurbished programme carries the same one-year limited warranty as a new Mac, is eligible for AppleCare, and replaces anything that failed testing with genuine Apple parts. One caveat worth knowing, because almost every article gets it wrong: the new battery and outer shell Apple advertises are guaranteed on refurbished iPhones and iPads, not on Macs. A refurbished Mac is tested and repaired, not rebuilt, so battery cycle count is worth asking about. The other catch is availability: stock appears and disappears within hours, and specific configurations, particularly the higher-RAM ones you actually want, are the first to go.

That is the problem RefurbMe exists to solve. We track refurbished Apple stock across Apple, Amazon Renewed, Back Market, eBay Refurbished and others in seven markets, so you can compare what is actually in stock right now rather than refreshing one store's page. If you want to see what a 16 GB Air is going for today, start with refurbished MacBooks.

Frequently Asked Questions

Does Claude Code need an Apple Silicon Mac? No. It runs on Intel Macs on macOS 13 or later. But the rest of your toolchain in 2026 increasingly assumes Apple Silicon, some tools no longer ship Intel binaries at all, and the battery life difference is enormous. If you are buying rather than using what you have, buy Apple Silicon.

Is 8 GB genuinely enough? For one session on one project, yes. For the way most people actually use an agent, which is several sessions in parallel, no. Treat 16 GB as the floor.

Does storage size affect Claude Code? Only indirectly. The agent reads and writes small files. But a nearly full SSD leaves no room for swap, and swap is what saves you when memory runs short. Keep 15 to 20 percent free, and if you are choosing, prefer more RAM over more storage. You can attach an external drive. You cannot attach RAM.

Should I buy a Mac mini instead and work remotely? It is a good setup if you already have a monitor and a keyboard, and it is the cheapest route to 24 GB. It is also the machine of choice for anyone running an always-on agent rather than an interactive session.

Does any of this change if I run models locally? Completely. Local inference is bound by memory capacity and bandwidth, which pushes you toward Pro, Max and Ultra chips and toward 32 GB and above. That is a different purchase for a different reason, and worth reading up on separately before you spend the money.

How do I read the Markdown files the agent writes? That part does not depend on which Mac you buy. macOS opens .md files as raw text in TextEdit on every machine, so plans, specs and READMEs arrive with their headings, tables, code blocks and Mermaid diagrams unrendered. MacMD Viewer ($19.99 one-time) is a native macOS reader that renders them properly, with a QuickLook extension so you can press Space on a .md in Finder without launching anything.