Claude Code Projects: what it is, where it runs, who has it
Claude Code's Projects beta: one coordinator conversation running parallel cloud threads. Plans, the waitlist, usage, memory, and how to start one.
Projects in Claude Code is one long-running conversation in which Claude acts as a coordinator: you send it a goal, bug reports or a list of tasks, and it starts a thread for each, a separate Claude Code session that works on its own branch, opens a pull request when the work calls for one, and reports back. Threads are cloud sessions on Anthropic’s machines unless you ask for one to run on your own computer through Remote Control, and the project keeps its instructions, a shared memory and a library of files between them. As of November 2026 it is a public beta on Pro, Max and Ultra, still rolling out (not on Team or Enterprise), used at claude.ai/code, in the desktop app’s Code tab and in the Claude mobile app, not in the terminal. Every running thread is a full session against your plan’s usage, and a new project runs threads on Opus at high effort until you change it.
What a project is
Anthropic’s Projects page describes it as “one ongoing conversation where Claude coordinates a stream of related work for you.” Two kinds of session make one up:
- The coordinator is the project conversation. It “decides what becomes a thread, and keeps track of every thread it started,” passes a new request to a thread already working in that area, and answers quick questions itself. “It sees what threads report back, not every step they take.” It has no connectors of its own.
- Threads do the work. Each is “a separate session with its own context window,” starts with the project’s instructions, works on a new branch from the repository’s default branch, and opens a pull request when you ask or on its own for a concrete fix. With auto-fix on, it watches that pull request and pushes fixes when CI fails or a reviewer comments. Its full results stay in the thread; you open it to read them, or to type to it directly.
An Overview pane beside the conversation sorts the threads into Waiting on you (a question, an approval, or a failure), Ready for review, Landing and Resolved, and a thread resolves by itself after a week with no activity. The docs suit it to work that outlasts one session and keeps producing tasks: one change across many repositories, a thread per repository, or an area you keep feeding with bug reports and stack traces.
Where the threads run
“Each thread is usually a cloud session”: a fresh Anthropic-hosted machine with the repository cloned, the same environment described in Claude Code on the web against your own machine. “Cloud threads keep going after you close your laptop,” and you can check on them and steer them from your phone. The environment, plugins and connectors a cloud thread gets are set on the project; it reads MCP tools from the connectors on your claude.ai account, and does not read permission rules from a repository’s .claude/settings.json.
At launch on September 17 that was the only option (“Threads run in the cloud today,” said Anthropic’s announcement). The docs now describe a thread on your own computer, for a task that needs its files, tools, MCP servers or Claude Code settings. It needs Claude Code v2.1.280 or later on that computer and the folder connected through Remote Control, either in the desktop app (Settings > Code, or Settings > Claude Code, then Use this computer from your phone and claude.ai) or from a terminal in that folder:
claude remote-control
# in a git repository, a worktree per thread:
claude remote-control --spawn worktreeCopiedYou then pick Work locally from the + menu (or Work in a local folder in the desktop app) and allow the card Claude sends. Three differences from a cloud thread are in the docs: it starts with the project’s instructions but “not with its memory files loaded”; it “pauses while that computer is asleep, and stops if the desktop app or claude remote-control quits”; and “two threads working in one folder at the same time can overwrite each other’s changes,” which is what --spawn worktree is for. On a Linux server the terminal form is the one that applies, and left running inside tmux it keeps the folder reachable after your SSH connection drops; the machine has to stay on for the thread to move.
Who has it, and the waitlist
“Projects are in public beta on Pro, Max, and Ultra plans and rolling out gradually,” the docs say, and “Projects aren’t offered on Team or Enterprise plans yet.” The rollout began with accounts that had used cloud sessions and had no existing projects in claude.ai chat or Cowork. To check yours, look for Projects in the sidebar at claude.ai/code or in the desktop app’s Code tab, then under More; if it is in neither, the rollout has not reached you and the waitlist is the way in. On October 9 Anthropic’s ClaudeDevs account posted “We just let in every Pro and Max user from the Claude Code Projects waitlist!” and that more would follow “as capacity allows”, so joining now is still a wait, of a length Anthropic does not give.
It is not available in the terminal CLI, the VS Code extension or the JetBrains plugin, nor through Amazon Bedrock, Google Cloud’s Agent Platform or Microsoft Foundry; there is no command to create or list projects from a shell.
Starting one
- At
claude.ai/codeor in the desktop app’s Code tab, choose Projects, then New project. If the button reads Get started, it first asks you to upgrade or archive the claude.ai projects it lists. - Give it a name (the only required field), optionally a one-line goal under “What do you want to work on?”, and Context: the GitHub repositories, plus any files, folders or Google Drive folders threads should read.
- Click Create project, then open Project settings from the menu in the project header for the rest: Environment, Memory > Project instructions, General for models and effort, Plugins.
An existing cloud session can join one through Continue in project in its sidebar menu, or become one through Continue as new project. A session you started on your own machine cannot be added.
What stays between threads
- Instructions: up to 16,000 characters, sent to every new thread and to the coordinator, so a rule such as which branch to target is stated once. They are instructions, not enforced settings.
- Memory: notes on requirements, decisions and pitfalls, kept as files. Every cloud thread reads the index,
MEMORY.md, when it starts; you ask Claude to remember something, or edit the files in Project settings > Memory. It is separate from theCLAUDE.mdfiles in your repositories, which load as usual. - The library: files you add and artifacts Claude produces. Uploads are copies, so a later change on your computer does not reach the project.
What does not reliably stay is uncommitted work. A cloud thread’s sandbox pauses between turns, and “if the sandbox can’t be resumed, the thread continues from a fresh clone, so uncommitted changes can be lost”; the docs’ advice on long tasks is to have Claude commit and push work in progress. Deleting a project removes its threads, memory and files for good, and leaves the branches and pull requests already on GitHub.
How it counts against your usage
“Project usage counts against the same plan limits as your other Claude Code sessions,” and each running thread “is a full session, and several can run at once.” Four threads running for an hour draw roughly what four sessions of your own would. The figures that govern it:
- The one enforced cap is 500 new threads a day across your projects. There is no fixed number running at once; Claude starts as many as the work calls for, and a limit you give it is “a preference rather than a cap.”
- A new project runs every thread on Opus at high effort, “which draws on your plan fastest.” Thread model and effort, and the coordinator’s, are separate settings in Project settings > General, and one task can ask for another model.
- Project settings > Usage shows tokens by thread and by model. A project with nothing running, no watched pull requests and no new messages uses nothing while it sits.
- Work goes past your plan’s limits only if usage credits are turned on for your account.
How the plan’s five-hour and weekly limits themselves work is in Claude Code’s usage limits.
Against the older Projects, agent teams and subagents
- Projects in claude.ai chat and Cowork (the version from 2024) are workspaces with their own chat history and knowledge base, “without threads or a coordinator.” They keep working as before; Anthropic’s help article says it has begun moving some Pro, Max and Ultra users’ existing projects to the new version, emails before moving one used in the last 30 days, and that “nothing is lost or deleted in the move.”
- Agent teams are “one session that starts teammate sessions for a single task,” in your terminal, and end with the lead session. A project outlives any one task and runs in the cloud.
- Subagents run inside one session, do a side task in their own context window and return a summary. A project’s threads can use subagents themselves.
- Agent view tracks several local sessions side by side, but “you still start each one and give it its task yourself.”
The limits it has in beta
- A project belongs to one user and cannot be shared, and there are no organization-level controls during the beta.
- A thread cannot move to another project, and two projects cannot be merged.
- Each thread has its own branch and copy of the repository, and the launch announcement says threads that touch the same code are reconciled as a merge conflict between pull requests, like any other; the docs describe no step of the coordinator’s for it.
- Anthropic gives no date for general availability or for Team and Enterprise.
If what you want is several long-lived sessions on a machine of your own, with your own tools and services always running and no cloud sandbox to resume, one arrangement is an orchestrator session in tmux on a computer that stays on, which works on any plan that includes Claude Code and needs no waitlist.