# Claude Code Projects: Is It Free? Plan Access and Beta Explained

> Source: <https://www.mindstudio.ai/blog/claude-code-projects-beta/>
> Published: 2026-11-18 00:00:00+00:00

**Key takeaway:** Claude Code Projects, in beta on Pro, Max, and Ultra plans, splits work between a planning coordinator model and execution models like Opus 5.5, replacing manual all-caps instructions and per-command approvals with auto mode, deny lists, sandboxing, and hooks.

# Claude Code Projects: Is It Free? Plan Access and Beta Explained

Claude Code Projects is in beta for Pro, Max, and Ultra plans. Here's how the multi-session orchestrator works and who can access it.

## What is Claude Code Projects?

Claude Code Projects is a new orchestration feature from Anthropic, currently in beta, that lets one main Claude session manage a whole list of tasks by spinning up separate sessions for each one and checking the results as they come back. Instead of opening a dozen chats and babysitting each one, you talk to a single coordinator session that delegates work, reviews what comes back, and reports to you. It’s not free in the sense of being a standalone product. It’s included as a beta feature on paid Claude plans, specifically Pro, Max, and Ultra.

## TL;DR

- **Claude Code Projects is in beta** and available to anyone on a Pro, Max, or Ultra Claude plan, not a separate paid add-on.
- The feature works like a **manager-and-workers setup** : one orchestrator session takes your full task list and launches individual sessions to handle each item.
- You can assign **different models to different roles** , using a strong reasoning model as the coordinator and a model suited for execution, like Opus 5.5, for the actual sub-tasks.
- The shift reflects a broader change in how Claude Code is used: Anthropic’s own internal code change volume reportedly grew roughly tenfold between May and September, pushing the team toward managing many parallel sessions instead of one at a time.
- Several manual habits that used to define Claude Code use, like writing all-caps “IMPORTANT” instructions or manually approving every command, are being replaced by built-in systems like auto mode, deny lists, sandboxing, and hooks.
- Projects doesn’t require coding tasks specifically. It can run non-code workflows too, like client reports, research briefs, or batches of content ideas.

*Enterprises use Remy.*

## Which Claude plans include Claude Code Projects?

Based on how Anthropic has rolled this out, Claude Code Projects is available in beta to users on the **Pro**, **Max**, and **Ultra** plans. It is not a separate purchase or a standalone product tier. It sits inside the existing Claude Code experience as a beta capability, meaning it’s still being refined and may change before a general release. If you already have one of these subscriptions, you have access to try it rather than needing to upgrade specifically for this feature.

This matters because Anthropic has historically gated some advanced agentic features behind usage limits tied to plan tier. With Projects still labeled as beta, expect behavior, limits, and interface details to shift as Anthropic collects more usage data and feedback, similar to how other beta agent features have evolved before reaching general availability.

## How does the orchestrator model work?

The core idea behind Claude Code Projects is a split between a coordinator and the workers it manages. You give the main session your full list of tasks, whether that’s a sprint’s worth of coding tickets, a batch of research assignments, or a set of client deliverables. The main session doesn’t do the work itself. Instead, it breaks the list into pieces, starts a separate Claude Code session for each piece, and checks what comes back against the goal you described.

This is conceptually similar to how sub-agents already work in Claude Code, where a parent agent delegates to child agents, but Projects applies that pattern at the level of an entire project rather than a single task. One useful detail: you can configure different models for different roles inside your project settings. The orchestrator benefits from a model that’s strong at planning and judgment, since its job is deciding what to delegate and evaluating the quality of what comes back. The worker threads benefit from a model tuned for raw execution on the actual coding or writing task. Anthropic’s newer models were cited as fitting these roles, with a reasoning-focused model coordinating and Opus 5.5 handling the thread-level work.

The practical effect is that you stop managing twelve open chats and instead manage one conversation, with the orchestrator reporting status, flagging what needs your attention, and surfacing problems before they pile up.

## Why is Claude Code shifting toward orchestration instead of single sessions?

The change tracks with how usage patterns have evolved inside Anthropic itself. The volume of code changes produced using Claude Code reportedly grew from around 2,500 per day in May to more than 30,000 per day by September, a scale that stopped being manageable through one-at-a-time supervision. At that volume, individuals on Claude Code teams effectively became managers of many concurrent sessions rather than operators typing prompts into a single chat window.

Build a vendor-risk app for our 300 buyers.

That shift exposed a problem with older best practices. Writing exhaustive instructions in a cloud.md file, loading it with all-caps directives like “IMPORTANT” or “NEVER,” made sense when models needed heavy steering. On stronger models, that kind of over-specification reportedly backfires, producing more tool calls and worse outcomes because the model gets pulled off track by instructions it didn’t need. The practical test that’s emerged: only include something in your configuration file if Claude genuinely couldn’t infer it by reading your code or project structure. Anything Claude can already figure out on its own is just noise that risks confusing a capable model.

## What replaced manual approval and oversight?

A second major shift involves how command approval works. The old pattern was binary: click “allow” on every single action Claude wanted to take, or disable permission checks entirely and hope nothing went wrong. Claude Code now defaults to an “auto mode” in both the terminal and desktop app, where a separate classifier evaluates each command and decides whether it’s safe to run automatically or needs a human to weigh in. Auto mode trusts actions inside your current folder and repository by default, and blocks things like pushes to unrelated deploy repositories until you explicitly approve them.

Anthropic’s internal testing reportedly found that human reviewers caught only around 14% of clearly dangerous commands when manually approving actions, while auto mode caught about 89% of them, a meaningfully higher catch rate, though not perfect, since some risky commands still got through.

For anything irreversible, like touching environment files containing API keys or deleting unsaved work, the recommended approach is a hard-coded deny rule rather than a prompt-level instruction. Deny rules sit below auto mode and get checked before the classifier even runs. For higher-stakes or long-running work, sandboxing restricts which files and network destinations a command can reach, and cloud-based sessions isolate the work entirely on a disposable machine, so even a compromised script can’t access local credentials.

## What should you check before trusting an orchestrated run?

Delegating a task to an autonomous session doesn’t mean skipping verification, it means moving verification earlier. The same prompt can produce very different outcomes depending on how a model interprets edge cases, so the recommended practice is building in checks along the way rather than only reviewing the final output. These checks can be automated (a test suite, a formatting check, a screenshot comparison) or more subjective (a description of what “done” should look like that Claude evaluates against its own output).

For checks that need to run every time, Claude Code supports “hooks,” small scripts triggered automatically after specific actions, like running tests after every edit or a full suite before a task is marked complete. The exit code these scripts return matters: scripts need to exit with a code that explicitly sends Claude back to fix the issue, otherwise a failed test can get silently marked as completed.

## Frequently Asked Questions

### Is Claude Code Projects free to use?

It isn’t a separate free product. It’s a beta feature included with paid Claude plans, Pro, Max, and Ultra, so access depends on already having one of those subscriptions.

### Do I need a coding background to use Claude Code Projects?

No. While it’s built on Claude Code, the orchestration pattern applies to non-coding workflows too, including research tasks, client reports, and content planning.

### Can I choose which model acts as the orchestrator?

Yes. Project settings let you assign a coordinator model separate from the model used in individual worker threads, allowing you to pair a planning-focused model with an execution-focused one.

*Nothing gets lost.*

**312 apps** across 41 teams · all in one place

### What happens if Claude runs a dangerous command?

Auto mode uses a classifier to catch risky actions automatically, but it isn’t perfect. For truly irreversible actions, like reading credential files, a hard deny rule is recommended since it blocks the action before auto mode even evaluates it.

### How is this different from sub-agents in Claude Code?

Sub-agents already let one session delegate to others for a single task. Claude Code Projects applies that same delegation pattern across an entire project or task list, with one orchestrator session managing multiple independent threads at once.
