Claude Code Projects: a coordinator and parallel threads

On November 14, 2026, Claude Code Projects hit the front page of Hacker News, with Boris Cherny from the Claude Code team answering questions in the thread. The feature itself is older: Anthropic announced it on September 17, 2026 as a beta for select accounts, and according to the documentation it is now a public beta on the Pro, Max and Ultra plans.

What are Claude Code Projects?

A project is one long conversation in which Claude acts as a coordinator. You paste in a bug report, a stack trace or a list of tasks, and for each one the coordinator starts a separate thread or hands it to a thread already working in that area. Short questions it answers on the spot.

Each thread is usually a full Claude Code session in the cloud, with its own context window and its own branch. When the task calls for it, the thread opens a pull request and then looks after it: it pushes fixes when CI fails and replies to review comments. Threads run in parallel and keep going after you close your laptop, and you can steer them from your phone too.

The Overview panel shows the state of everything, with threads grouped into “Ready for review”, “Waiting on you”, “Working” and so on. The coordinator only sees what the threads report back to it (for the details you have to open the thread itself).

Cherny wrote on HN that nearly all of his code gets written this way now: “Projects are how I write pretty much all my code these days”. By his account the coordinator is there to fix the problem of attention scattered across many sessions, though he added that not every job needs to be parallelized.

What has changed since the September beta?

First, threads on your own computer. In September threads ran only in the cloud, and a local version was coming “very soon”. The documentation now describes it: you pick Work locally in the project chat, and the thread becomes a Claude Code session in a folder on your machine, connected through Remote Control (you need Claude Code 2.1.280 or later). A thread like that uses your local files, tools, MCP servers and hooks, but it only runs while the computer is awake.

Second, availability. The September beta covered select Pro and Max subscribers. Now it is a public beta on Pro, Max and Ultra, still rolling out gradually (if you don’t see Projects in the sidebar or under “More”, there is a waitlist). Team and Enterprise plans are still not supported.

The HN thread links to an AlphaSignal piece that also lists sharing threads by link and reminders, among other things. The documentation still says a thread can’t be shared with another person, so I’m leaving those out.

How fast does a project burn through your usage limit?

Faster than a single session. A project draws on the same plan limits as the rest of Claude Code, and every running thread is a full session. The documentation says that on the Pro plan you will hit your limit sooner on days when you work with a project.

A new project runs Opus everywhere: threads at high effort (the setting that, per the docs, uses up your plan the fastest) and the coordinator at low. There is no fixed number of threads running at once. If you ask for “at most two at a time”, Claude treats that as a preference, not a hard block. The only enforced cap is 500 new threads per day across all projects.

On HN, Cherny said that in Anthropic’s A/B tests and evals the coordinator adds less than 10% in tokens compared with launching the sessions yourself. He also admitted the team doesn’t yet have a good way to budget usage, and that for now putting a budget in the prompt, such as “use around 300k tokens”, works to some extent. Cost was exactly what commenters worried about: one user described seven subagents on Fable 5.2 eating through their Max 5x limit in about 10 minutes.

What you can do about it:

1. Change Thread model and Thread effort under Project settings > General to a smaller model or lower effort.
2. Start with one small task and see how the thread reports back before you send the whole list.
3. Check the Usage tab, which shows token usage by thread and by model.

A thread that stops at the limit waits by default and picks up on its own once the limit resets. There is more on what eats your limit in Claude Code in the post on background agent statuses.

What should you watch out for?

First, what cloud threads don’t load from the repository. They get CLAUDE.md, but not the skills in .claude/skills/, the MCP servers in .mcp.json, or the hooks and permission rules in .claude/settings.json. Skills and connectors have to be enabled on your claude.ai account, and plugins added in the project settings.

Second, uncommitted changes. A thread’s sandbox is paused between turns, and if it can’t be resumed, the thread starts over from a fresh clone. For long tasks the documentation advises asking Claude to commit and push as it goes.

Third, the requirements. The code has to live on github.com (not GitHub Enterprise Server, GitLab or Bitbucket), with the Claude GitHub App installed. A project belongs to one person, and for now there are no organization-level settings at all.

FAQ

Do Claude Code Projects work in the terminal?

No. Projects are available on claude.ai/code, in the desktop app and in the Claude mobile app. They are not in the CLI, the VS Code extension or the JetBrains plugin. They don’t work through Amazon Bedrock, Google Cloud or Microsoft Foundry either. If you are just getting started with the tool itself, see the Claude Code tutorial.

Do Claude Code Projects cost extra?

The documentation doesn’t list a separate fee. A project uses your plan’s limits and can’t go over them on its own. Working past the limit is only possible if you have usage credits turned on for your account.

Does a project need a repository?

No. Cloud threads can do research, write documents and run code in their own sandbox, and the finished files land in the Library tab. Instead of a repository you can upload files, folders or Google Drive folders.