Checked against Anthropic's own docs, September 2026
Claude Code CLI: Install Claude Code, Terminal Commands, Pricing and Usage Limits
Eight install routes, and only one of them keeps itself up to date. Before any of that, the fact most write-ups still get wrong: the free Claude.ai plan does not include Claude Code at all.
No card to start · Flat $29/mo billed yearly, no usage meter to watch
Pick a task
Plan
- planning
Files changed
Test run
Pull request
You review and merge. Agentcode never merges on its own.
The short answer
Claude Code CLI is Anthropic's terminal coding agent, and it needs a paid account: the free Claude.ai plan does not include it, so you need Pro, Max, Team, Enterprise or a Claude Console account. Install it with curl -fsSL https://claude.ai/install.sh | bash on macOS, Linux or WSL, irm https://claude.ai/install.ps1 | iex on Windows, brew install --cask claude-code, winget install Anthropic.ClaudeCode, a signed apt, dnf or apk repository, or npm install -g @anthropic-ai/claude-code. Then run claude in a project directory. Only the native install updates itself in the background; every package manager route has to be upgraded by hand.
How to install Claude Code
Every command below is copied from Anthropic's own setup documentation. Pick the route that matches how the rest of your developer tooling is managed, not the one printed first, because the route you choose decides something that is not obvious at install time: whether the agent will keep itself current. That is the next table, and it is the one worth reading twice.
Swipe to see more
| Route | Command | What to know |
|---|---|---|
| Native installer (macOS, Linux, WSL) | curl -fsSL https://claude.ai/install.sh | bash | The route Anthropic lists first and the only one that keeps itself current. It installs a standalone binary, so the machine does not need Node.js at all. Add "-s stable" to take the slower release channel, or "-s 2.1.89" to pin an exact version. |
| Native installer (Windows PowerShell) | irm https://claude.ai/install.ps1 | iex | Native Windows, no WSL and no Administrator rights needed. Your prompt shows "PS C:\" in PowerShell. Installing Git for Windows is optional but recommended, because it gives Claude Code a real Bash tool instead of falling back to PowerShell. |
| Native installer (Windows CMD) | curl -fsSL https://claude.ai/install.cmd -o install.cmd && install.cmd && del install.cmd | Use this one when your prompt reads "C:\" without the PS. If you see "The token '&&' is not a valid statement separator" you are in PowerShell and want the line above instead. |
| Homebrew (macOS, Linux) | brew install --cask claude-code | A cask, and there are two of them. claude-code follows the stable channel, roughly a week behind. claude-code@latest ships every release immediately. On Homebrew the channel is chosen by which cask you install, not by a setting. Upgrade with brew upgrade claude-code. |
| WinGet (Windows) | winget install Anthropic.ClaudeCode | Convenient if your fleet is already managed through WinGet. It does not auto-update: run winget upgrade Anthropic.ClaudeCode. The upgrade can fail while Claude Code is running, because Windows locks the executable. |
| apt (Debian, Ubuntu) | sudo apt update && sudo apt install claude-code | Anthropic publishes a signed apt repository with stable and latest channels. Register the signing key first and check the fingerprint reads 31DDDE24DDFAB679F42D7BD2BAA929FF1A7ECACE. Skip the key and apt update later fails with NO_PUBKEY BAA929FF1A7ECACE. |
| dnf (Fedora, RHEL) and apk (Alpine) | sudo dnf install claude-code | Same signed-repository model. dnf prompts you to confirm the same fingerprint on first install. On Alpine you also need bash, curl, libgcc, libstdc++ and ripgrep, and you have to set USE_BUILTIN_RIPGREP to 0 in settings.json. |
| npm (any platform) | npm install -g @anthropic-ai/claude-code | Installs the identical native binary through a per-platform optional dependency, so the running agent never touches your Node.js. Node 22 or later since v2.1.198. Upgrade with npm install -g @anthropic-ai/claude-code@latest, not npm update -g. Never use sudo npm install -g. |
| Check it worked | claude --version | A healthy install prints a version like 2.1.211 (Claude Code). For a fuller read, claude doctor prints install health, settings-file validation errors and the result of the last update attempt, all without starting a session. |
After it installs, open a terminal in the project you want to work in and run claude. The first launch prompts you to log in through the browser. If ANTHROPIC_API_KEY is already set in your environment, Claude Code asks once to approve that key instead of opening a browser at all, which is a detail worth knowing before you wonder why a machine signed itself in to the wrong account.
Does Claude Code CLI auto update?
Only if you installed it natively. This is the single most useful thing on this page for anyone rolling the CLI out to more than one person, and it is buried in a note in Anthropic's docs rather than stated as a comparison. The install route you standardize on quietly decides whether your developers stay current or drift onto builds from three months ago.
Swipe to see more
| Install route | Auto-updates? | Manual upgrade command | What that means in practice |
|---|---|---|---|
| Native installer | Yes | claude update | Checks on startup and periodically while running, downloads in the background and applies the new version next launch. This is the only route where a developer who never thinks about updates stays current. |
| Homebrew | No, opt-in | brew upgrade claude-code | Set CLAUDE_CODE_PACKAGE_MANAGER_AUTO_UPDATE to 1 and Claude Code will run the upgrade for you in the background and prompt for a restart. Homebrew keeps old versions on disk, so run brew cleanup occasionally. |
| WinGet | No, opt-in | winget upgrade Anthropic.ClaudeCode | Same opt-in environment variable as Homebrew. Windows locks the running executable, so if the upgrade fails Claude Code shows you the manual command rather than retrying behind your back. |
| apt, dnf, apk | No, and cannot be | sudo apt upgrade claude-code | These stay manual because the upgrade needs elevated privileges. Updates arrive through your normal system patching workflow instead, which for most regulated teams is the point rather than a drawback. |
| npm | Yes, with a caveat | npm install -g @anthropic-ai/claude-code@latest | It self-updates unless the npm global directory is not writable, in which case you get a one-time startup notice and claude doctor lists the fixes. Avoid npm update -g: it honors the semver range from the original install and may leave you behind. |
The asymmetry is deliberate on Anthropic's side. Package managers own their own upgrade lifecycle, and apt, dnf and apk need elevated privileges that a user-level agent has no business acquiring on its own. But the practical result is that a team which standardizes on Homebrew because it standardizes on Homebrew for everything has, without discussing it, opted out of updates. Setting CLAUDE_CODE_PACKAGE_MANAGER_AUTO_UPDATE to 1 puts that back, and it is a one-line entry in a settings file that ought to be part of any rollout checklist.
Is Claude Code CLI free? What plan do you actually need?
It is not free, and the way it is not free surprises people. Anthropic states in three separate documents that Claude Code requires a Pro, Max, Team, Enterprise or Claude Console account, and that the free Claude.ai plan does not include Claude Code access. There is no reduced free tier of the CLI. Its closest rival goes the other way: Codex CLI does run on ChatGPT Free, so if a colleague told you the terminal agent was free, they were probably talking about a different one.
Swipe to see more
| Account | CLI included | Starting permission mode | What it means for terminal work |
|---|---|---|---|
| Claude Free | No | Not applicable | This is the fact most write-ups get wrong. Anthropic states plainly that the free Claude.ai plan does not include Claude Code access. There is no free tier of the CLI, no trial allowance and no reduced-capability mode. Codex CLI runs on ChatGPT Free; Claude Code does not have an equivalent. |
| Pro | Yes | Auto | The entry paid plan and where most individual developers start. Interactive terminal sessions start in auto mode, where a classifier reviews actions instead of prompting you for each one. Chat and coding draw on the same allowance, so heavy Claude chat use eats the budget you had for sessions. |
| Max | Yes | Auto | Sold in two multiples of Pro. This is where developers who run the agent all day end up, because agentic turns carry file contents, tool results and multi-step reasoning rather than a few paragraphs of chat. |
| Team | Yes | Auto | Adds central billing, admin settings and org analytics, with standard and premium seats that can be mixed so the few engineers who live in the agent get the larger allowance. Managed settings become available, which is where version pinning and policy enforcement live. |
| Enterprise | Yes | Manual | Note the mode column. Anthropic lists auto mode as the built-in starting mode on Pro, Max and Team, and says other plans start in Manual. So an Enterprise seat asks before acting where a personal Pro seat does not, out of the box. Adds SSO, audit logs and per-member spend limits. |
| Claude Console (API) | Yes | Manual | Pay per token instead of drawing on a subscription. No usage window to wait out, and no ceiling either until you set a workspace spend limit. This is the sign-in most teams use for CI, and it is the one where --max-budget-usd earns its keep. |
| Bedrock, Google Cloud, Microsoft Foundry | Yes | Manual | Claude Code can authenticate against a third-party provider instead of Anthropic, which is how teams that already have a cloud commit or a data-residency requirement run it. Billing and model availability follow that provider, not your Claude plan. |
Read the permission-mode column before you read the price column. Anthropic documents auto mode, in which a classifier reviews the agent's actions instead of prompting you for each one, as the built-in starting mode for interactive terminal sessions on Pro, Max and Team plans, and says other plans start in manual mode. That means the same claude command behaves more conservatively on an Enterprise or Console seat than on a personal Pro seat, out of the box, with no setting changed. It is a product difference hiding inside a plan name, and it is the sort of thing that makes a pilot on personal seats feel faster than the rollout that follows it. Plan-by-plan prices sit on our Claude Code pricing breakdown.
Claude Code terminal commands worth memorizing
The full reference runs to hundreds of flags, and Anthropic notes that claude --help does not list every one of them, so a flag missing from help is not a flag that does not exist. These are the commands that carry the everyday work.
Swipe to see more
| Command | What it does |
|---|---|
| claude | Start an interactive session in the current directory. This is the whole product most days. |
| claude "query" | Start interactive with an opening prompt, so you skip a round trip. |
| claude -p "query" | Print mode: run the query through the Agent SDK, print the result and exit. This is the CI entry point and the one every flag below attaches to. |
| cat file | claude -p "query" | Pipe content in. Useful for feeding a failing test log or a diff straight into the agent from a shell script. |
| claude -c | Continue the most recent conversation in this directory. Add -p to continue non-interactively. |
| claude -r "<session>" "query" | Resume a specific session by ID or by the name you gave it. |
| claude update | Apply an update immediately rather than waiting for the next background check. Reports the old and new version, or that you are already current. |
| claude install [version] | Install or reinstall the native binary. Accepts a version like 2.1.118, or the channel names stable and latest. |
| claude doctor | Read-only diagnostics without starting a session: install health, settings validation errors, and the outcome of the last update attempt. Run this before filing a bug. |
| claude auth status | Prints auth status as JSON, or human-readable with --text. Exits 0 when signed in and 1 when not, which makes it the clean way to gate a CI step. |
| claude setup-token | Generate a long-lived OAuth token for CI and scripts. It prints the token and does not save it. Requires a Claude subscription. |
| claude mcp | Configure Model Context Protocol servers. claude mcp login <name> runs a server OAuth flow without opening the interactive panel. |
| claude agents | Open the agent view to monitor and dispatch parallel background sessions. claude attach, claude logs, claude stop and claude respawn drive the same sessions from the shell. |
| claude project purge [path] | Delete all local state for a project: transcripts, task lists, debug logs, file-edit history and prompt history. The command to know before you hand a laptop back. |
Two of those deserve more attention than they get. claude auth status exits 0 when signed in and 1 when not, which is exactly what you want at the top of a pipeline step rather than discovering the credential expired halfway through a run. And claude doctor reports settings-file validation errors, which is where a surprising share of "it just stopped working" turns out to live.
Running Claude Code CLI headless, and how to cap what it spends
Print mode, claude -p, is how the CLI runs unattended. It is also where Claude Code has a genuine advantage that almost nobody writes about, and we went looking for the counterexample before saying so: we read OpenAI's codex exec documentation on the same day and it publishes no per-run dollar cap and no turn cap at all.
Swipe to see more
| Flag | What it does | Why it matters in a pipeline |
|---|---|---|
| --max-budget-usd | A hard dollar ceiling for a single print-mode run. Anthropic documents that subagent spend counts toward the same cap and that spawning another subagent fails once the cap is reached. | This is the flag that makes an agent safe to put behind a webhook. Codex CLI has no published equivalent, and neither does Cursor CLI: both meter you after the fact rather than letting you name the number first. |
| --max-turns | Cap the number of agentic turns. Exits with an error when the limit is reached. No limit by default. | A dollar cap stops runaway spend; a turn cap stops a runaway loop before it becomes spend. Use both. The default of no limit is the one people discover the expensive way. |
| --output-format json | Structured output from print mode. stream-json emits events as they happen, and --include-partial-messages adds partial streaming events. | Parse the result in a shell step instead of scraping text. --json-schema goes further and validates the final output against a schema you supply. |
| --bare | Skip auto-discovery of hooks, skills, custom commands, subagents, plugins, MCP servers, auto memory and CLAUDE.md so scripted calls start faster. | A CI runner has none of your local configuration anyway. Skipping the discovery pass removes a class of "works on my machine" difference between a developer laptop and the pipeline. |
| --permission-mode | Start in default, acceptEdits, plan, auto, dontAsk, bypassPermissions or manual. --dangerously-skip-permissions is the same as bypassPermissions. | Set this explicitly in CI rather than relying on the default, because the default is not the same on every plan: Pro, Max and Team start in auto, other plans start in manual. |
| --model and --effort | Pick the model by alias (sonnet, opus, haiku, fable) or full name, and set effort to low, medium, high, xhigh, max or ultracode. | Effort is the cheaper lever. Dropping a routine lint-fix job from high to low changes the cost of every run in the pipeline without changing the model you validated against. |
| --fallback-model | A comma-separated list of models to try in order when the primary is overloaded or has been retired. | A pipeline that fails because a model was busy is a pipeline people learn to ignore. This is a two-word fix for a class of flaky failure. |
The difference is worth stating plainly, because it changes how you can safely automate. Codex CLI meters a rolling five-hour message window and stops you when it is spent, which protects OpenAI's capacity rather than your budget. Cursor CLI bills monthly pools at model rates and tells you afterwards. Claude Code lets you name the number in dollars before the run starts and refuses to exceed it, subagents included. If you are wiring an agent to a webhook or a nightly job, that is the difference between a budget and a hope. We put the three side by side in what Copilot CLI, Claude Code and Codex CLI actually cost, and covered the Codex side in detail on the Codex CLI page.
Pinning versions, release channels and a verifiable supply chain
If a security review stands between your team and this tool, this section is the one to send them. Anthropic ships more change-control machinery than the tool's reputation as a developer toy suggests, and two of these settings are routinely confused with each other in a way that produces a guarantee nobody actually has.
Swipe to see more
| Control | Effect | Detail |
|---|---|---|
| autoUpdatesChannel | latest (default) or stable | The stable channel serves a build roughly one week old and skips releases with major regressions. On Homebrew this is chosen by cask name instead: claude-code is stable, claude-code@latest is latest. |
| minimumVersion | Floor for updates only | Background updates and claude update refuse to install anything below this. Moving to the stable channel therefore will not downgrade a developer who is already ahead of it. It does not stop an older build from running. |
| requiredMinimumVersion / requiredMaximumVersion | Claude Code refuses to start | Set in managed settings, these bound the versions that are allowed to run at all, not just to install. This is the pair you want if a security review needs a guarantee rather than a default. |
| DISABLE_AUTOUPDATER | Stops the background check | claude update and claude install still work. If you need to block every update path because you distribute the binary yourself, set DISABLE_UPDATES instead. |
| Signed packages | Verifiable supply chain | Every release publishes a manifest.json of SHA256 checksums signed with an Anthropic GPG key, so verifying one signature transitively verifies every platform binary. Signatures exist from 2.1.89 onward. macOS builds are signed by Anthropic PBC and notarized by Apple; Windows builds carry an Authenticode signature. |
The confusion to avoid: minimumVersion only constrains what an update is allowed to install. It will not stop an older build that is already on the machine from running. If your control needs to be a guarantee rather than a default, requiredMinimumVersion and requiredMaximumVersion in managed settings are the pair that makes Claude Code refuse to start outside the range you approved.
Claude Code CLI system requirements
- macOS 13.0 or later, Windows 10 1809 or later, Windows Server 2019 or later, Ubuntu 20.04 or later, Debian 10 or later, or Alpine Linux 3.19 or later.
- 4 GB or more of RAM, on an x64 or ARM64 processor.
- An internet connection. The agent runs locally but every model call goes out to Anthropic or to your configured provider.
- Bash, Zsh, PowerShell or CMD. On native Windows, installing Git for Windows is what gives Claude Code a real Bash tool instead of the PowerShell fallback.
- ripgrep, which normally ships with Claude Code. On Alpine and other musl distributions you install it yourself and set USE_BUILTIN_RIPGREP to 0.
- A location in Anthropic's list of supported countries, and a Pro, Max, Team, Enterprise or Console account.
Does Claude Code CLI work on Windows?
Yes, three different ways, and only one of them supports sandboxed command execution. Pick deliberately.
| Option | Sandbox | When to use it |
|---|---|---|
| Native Windows | Not supported | Windows-native projects and tools. Without Git for Windows the agent runs shell commands through the PowerShell tool instead of Bash. |
| WSL 2 | Supported | Linux toolchains, or any time you want sandboxed command execution. Install and launch claude inside the WSL terminal, not from PowerShell. |
| WSL 1 | Not supported | Only when WSL 2 is unavailable on the machine. |
How much does Claude Code cost per developer?
On a subscription, it is the seat price and nothing else. On Claude Console or a cloud provider it is metered, and Anthropic publishes what that has actually looked like across enterprise deployments. These are ranges, and we are quoting them as ranges rather than turning them into a tidy per-session figure that Anthropic does not publish.
Swipe to see more
| Measure | Published figure | How to read it |
|---|---|---|
| Average per developer, per active day | About $13 | Anthropic's own figure, observed across enterprise deployments. An "active day" is a day the developer actually used it, not a calendar day, which is why this number looks small next to the monthly one. |
| Average per developer, per month | $150 to $250 | Quoted as a range because it is one. Against a $20 Pro seat that is seven to twelve times the subscription price, which is the number that surprises finance teams moving from seats to API billing. |
| 90th percentile, per active day | Under $30 | Nine out of ten users stay below $30 on a day they use it. The useful read is that the tail is what costs money: a small number of heavy automation users move the average. |
| What drives the variance | Model, repo size, parallelism | Anthropic names model selection, codebase size and usage patterns such as running multiple instances or automation. Effort level is the lever most teams never touch and the cheapest one to pull. |
| How to see it | /usage in session | Shows plan usage bars, activity stats and a breakdown. The Session block reports API token cost and is aimed at Console users; on Pro and Max the subscription already covers it, so that figure is not a bill. |
The number that reframes a budget conversation is the middle row. A Pro seat is $20 a month. Anthropic's own observed range for metered usage is $150 to $250 per developer per month. That is seven to twelve times the seat price for the same person doing the same job, and the gap is not a markup, it is what agentic work actually consumes when nobody is watching a meter. It is also why the pilot always looks cheaper than the rollout: pilots run on personal subscriptions where the ceiling is enforced for you, and production runs on Console where it is not, until you enforce it yourself. We work through that arithmetic for a whole team on what an AI coding agent really costs per developer.
Which install route should your team standardize on?
One developer, own laptop
Native installer, latest channel, Pro seat. You get background updates you never have to think about and the newest agent behavior as it ships. This is the default for a reason and there is no good argument against it for an individual.
A team on macOS
Homebrew, and decide the channel deliberately by picking the cask: claude-code for stable, claude-code@latest for the newest. Then set CLAUDE_CODE_PACKAGE_MANAGER_AUTO_UPDATE to 1 so nobody drifts three months behind without noticing.
A managed Windows fleet
WinGet, so the package lands in the same inventory as everything else you ship. Expect the occasional upgrade to fail while the agent is running, because Windows locks the executable, and tell people the manual command rather than letting them assume it is broken.
A regulated or change-controlled team
The signed apt, dnf or apk repository on the stable channel, plus requiredMinimumVersion and requiredMaximumVersion in managed settings. That combination gives you a verifiable package source, a build that is deliberately a week old, and a version range outside which the agent will not start at all.
CI and automation
A Console account or a cloud provider rather than a seat, claude setup-token for the credential, claude auth status to gate the step, and print mode with --max-budget-usd, --max-turns, --bare and --output-format json. Set the permission mode explicitly instead of inheriting it.
A team already pinned through package.json
npm, but upgrade with npm install -g @anthropic-ai/claude-code@latest rather than npm update -g, which respects the original semver range and quietly leaves you behind. Node 22 or later, and never with sudo.
The fine print nobody puts in the quickstart
- The free Claude.ai plan does not include Claude Code. Every "get started free" write-up that says otherwise is out of date, and it is the single most common factual error about this tool.
- The default permission mode is not the same on every plan. Pro, Max and Team start interactive terminal sessions in auto mode, where a classifier approves actions for you. Other plans start in manual mode. Same command, different behavior, decided by your billing tier.
- Only the native install updates itself. Choosing Homebrew, WinGet or a Linux package manager is also choosing to run an update command, or to set CLAUDE_CODE_PACKAGE_MANAGER_AUTO_UPDATE where that is supported.
- minimumVersion and requiredMinimumVersion are not the same thing. The first only constrains what an update may install. The second, in managed settings, stops Claude Code running at all outside the range.
- npm update -g is the wrong upgrade command. It honors the semver range recorded at first install, so it can report success while leaving you on an old build. Use npm install -g @anthropic-ai/claude-code@latest.
- A missing apt signing key does not fail loudly at install time. It fails later, at apt update, with NO_PUBKEY BAA929FF1A7ECACE, by which point nobody remembers the step they skipped.
- Sandboxed command execution on Windows requires WSL 2. Native Windows and WSL 1 do not support it, which matters if your security posture assumed the sandbox was always there.
- There is no default limit on agentic turns in print mode. Anthropic documents "no limit by default", so an unattended run has no ceiling until you give it one with --max-turns and --max-budget-usd.
- On Pro and Max the session cost figure in /usage is not a bill. Those plans include the usage, so the dollar number shown there is aimed at Console users and is not what you owe.
Where we fit, and where Claude Code CLI beats us
The losses first. Claude Code CLI is an excellent terminal agent with the most serious change-control story of any tool in this category: signed packages, a stable release channel, version ranges outside which it will not start, and a real dollar cap on unattended runs. If your team wants an agent inside the terminal people already live in, buy it. Nothing we sell replaces that.
What we do differently is where the work happens. A CLI agent needs you in front of it, reading output and re-steering when it drifts, and it bills for every minute of that attention. Agentcode takes a task against a connected GitHub or GitLab repository, plans it, edits the code, runs your test suite and opens a pull request for a human to read. It never merges on its own, so your branch protection and your reviewers stay exactly where they are.
It also costs $29 a month, flat, billed yearly, which means there is no per-active-day figure to model and no metered range between $150 and $250 to explain to finance. That is a narrower product than Claude Code and we are glad to say so. If you are still building a shortlist, the AI coding tools worth shortlisting and every AI coding tool priced side by side cover the field honestly, including where we lose.
Claude Code CLI: the questions people actually search
Is Claude Code CLI free?
No. Anthropic states that Claude Code requires a Pro, Max, Team, Enterprise or Claude Console account, and that the free Claude.ai plan does not include Claude Code access. There is no free tier and no trial allowance in the CLI itself. If you want to try a terminal agent without paying, Codex CLI does run on ChatGPT Free, though OpenAI publishes no message limits for that tier.
Does Claude Code CLI require a subscription?
Yes, or a Console account. The accepted sign-ins are a Claude subscription (Pro, Max, Team or Enterprise), a Claude Console account billed per token, or a supported cloud provider such as Amazon Bedrock, Google Cloud or Microsoft Foundry. Console and the cloud providers bill usage rather than a seat, which is normally the right choice for CI.
How to install Claude Code on Windows
Run irm https://claude.ai/install.ps1 | iex in PowerShell, or the curl install.cmd line in CMD. Neither needs Administrator rights. You then choose between native Windows, WSL 2 and WSL 1: only WSL 2 supports sandboxed command execution. On native Windows, installing Git for Windows is optional but it is what gives the agent a real Bash tool.
How to install Claude Code on Mac
Either curl -fsSL https://claude.ai/install.sh | bash for the native installer, or brew install --cask claude-code if Homebrew already manages your tools. The native route auto-updates and the Homebrew route does not. macOS 13.0 or later is required, and the binaries are signed by Anthropic PBC and notarized by Apple.
How to install Claude Code on Ubuntu
The native installer works: curl -fsSL https://claude.ai/install.sh | bash on Ubuntu 20.04 or later. For a managed fleet, register Anthropic's signed apt repository instead and install claude-code with apt, which routes updates through your normal patching workflow. Check the signing key fingerprint reads 31DDDE24DDFAB679F42D7BD2BAA929FF1A7ECACE before you trust it.
Does Claude Code CLI auto update?
It depends entirely on how you installed it. Native installs check on startup and periodically while running, then apply the update on next launch. Homebrew and WinGet do not update automatically unless you set CLAUDE_CODE_PACKAGE_MANAGER_AUTO_UPDATE to 1. apt, dnf and apk never do, because those upgrades need elevated privileges. npm self-updates unless its global directory is read-only.
Can I use Claude Code CLI for free?
Not on the Claude free plan, which excludes Claude Code entirely. The cheapest legitimate routes in are a Pro subscription or a Claude Console account with a small workspace spend limit, which lets you pay a few dollars to evaluate it rather than committing to a month. A Console key also unlocks --max-budget-usd, so you can cap what an evaluation costs.
Does Claude Code CLI run locally?
The agent runs on your machine and reads, edits and executes against your local checkout, but the reasoning happens in Anthropic's models, so an internet connection is required. Nothing about it is offline. If your constraint is data residency rather than locality, authenticating through Amazon Bedrock, Google Cloud or Microsoft Foundry keeps the model calls inside a provider you already have an agreement with.
Does Claude Code CLI work on Windows?
Yes, natively on Windows 10 1809 or later and Windows Server 2019 or later, and also inside WSL. The practical difference is sandboxing: WSL 2 supports sandboxed command execution and native Windows and WSL 1 do not. Pick native Windows for Windows-native projects and WSL 2 for Linux toolchains or when you want the sandbox.
How much does Claude Code cost per developer?
On a subscription it is the seat price. On Console or a cloud provider it is usage, and Anthropic publishes what that has looked like in practice: about $13 per developer per active day and $150 to $250 per developer per month, with 90% of users staying under $30 on an active day. Treat those as ranges, because model choice, repo size and parallel sessions move them a long way.
What is the difference between claude update and claude install?
claude update applies the newest version on your current channel right away instead of waiting for the background check. claude install takes an argument, so it can install or reinstall a specific version such as 2.1.118, or switch you onto the stable or latest channel. Use update for routine currency and install when you need to pin or roll back.
Can I cap what Claude Code CLI spends in CI?
Yes, and this is where it separates from its rivals. Print mode accepts --max-budget-usd, a hard dollar ceiling for a single run, with subagent spend counting toward the same cap. Pair it with --max-turns to stop a loop before it becomes spend. OpenAI's codex exec documentation publishes no equivalent per-run dollar or turn cap.
Last updated: September 2026
Keep researching
More on Claude Code: Claude Code pricing plan by plan, Codex against Claude Code, Cursor against Claude Code and what Claude Code costs on Amazon Bedrock.
The other terminal agents: Codex CLI install, pricing and limits, GitHub Copilot CLI pricing and install, Gemini CLI against Claude Code and Cursor CLI against Codex CLI.
A terminal is not the only place an agent can work
Connect a GitHub or GitLab repository, describe a task, and read the pull request it opens. Flat $29/mo billed yearly, no metered range to model, and nothing merges without you.