Anthropic 把 Claude Code 的 Projects 改了：合上笔记本，云上的并行会话还在跑Anthropic Reworked Claude Code Projects: Close Your Laptop and the Parallel Cloud Sessions Keep Running

Anthropic reworked Projects in Claude Code. The old version of a project was just a folder: a pile of files plus a conversation. The new version is a single conversation that keeps moving forward, with Claude acting as the dispatcher. You explain what you want, and Claude decides which pieces of work should be split into a thread. Each thread is a full Claude Code cloud session running on its own branch and its own copy of the repo. They work in parallel, report results back to the main conversation, and they don't stop when you close your laptop. Anthropic's launch post describes it as "a conversation that splits itself into a bunch of parallel cloud sessions."

The old project was a folder — a pile of files plus a chat. Now it's a single conversation that keeps moving forward, with Claude acting as the dispatcher. Each thread is a full cloud session running its own branch and its own copy of the repo, working in parallel, and they keep going after you shut your computer.

The Concept of a Project Has Been Rewritten

Anthropic reworked Projects in Claude Code. The old version of a project was just a folder: a pile of files plus a conversation. The new version is a single conversation that keeps moving forward, with Claude acting as the dispatcher. You explain what you want, and Claude decides which pieces of work should be split into a thread. Each thread is a full Claude Code cloud session running on its own branch and its own copy of the repo. They work in parallel, report results back to the main conversation, and they don't stop when you close your laptop. Anthropic's launch post describes it as "a conversation that splits itself into a bunch of parallel cloud sessions."

Dispatch and Threads

The structure has two layers. The project conversation is the dispatch layer: it reads what you send, answers simple questions on the spot, and spins up threads when there's real work to do. It only sees the results threads report back — it doesn't watch every step they take. Threads are the layer that does the work: when a PR is needed they open one themselves, then babysit it with auto-fix — push a fix when CI fails, report back once checks pass.

Threads can break down further still. Each thread can use subagents, loops and workflows to split its own tasks out again.

Anthropic gives an example of its own: set the goal to "reduce p75 latency in the checkout flow," then have Claude profile each endpoint in turn, try optimizations, and open PRs in parallel threads. A second example is sunsetting a deprecated v1 endpoint across three repos at once — API, web and mobile — one thread per repo, with Claude reporting which PRs should merge first.

When two threads touch the same stretch of code, the overlap surfaces as an ordinary git merge conflict.

What Every Thread Inherits

Standing context is set once, and every new thread after that gets it: the project's repos and uploaded files, up to 16,000 characters of project instructions, and a project memory that Claude reads and writes through a MEMORY.md index. The memory holds things like: the release moved to Friday, or who to check with before touching billing.

Every thread also clones each repo in the project and loads its CLAUDE.md, skills and plugins. Permission rules, hooks and environment variables are different: they only apply in the directory where the thread was started, so single-repo projects honor them while multi-repo projects don't. MCP tools come from the connectors configured on your Claude account; the project conversation itself doesn't carry connectors, so any work that needs them has to go to a thread.

The Overview panel groups threads by status: awaiting review, needs your input, in progress, landing, idle, resolved. The Library tab holds uploaded files and files produced by threads.

The Cost

A project burns through plan quota faster than a single session, because every running thread is a full session. Anthropic provides a Usage tab broken down by project, and lets you set the model and effort level separately for the dispatch layer and for threads. New projects start with Opus everywhere: high effort for threads, low effort for the conversation. Idle threads also get woken when CI fails or someone leaves review comments, and keep burning quota. The hard cap is 200 new threads per day across all your projects; threads that hit a usage limit wait and then resume on their own.

Key Takeaways

• A project is one conversation, and Claude opens a thread for each piece of work
• Each thread is a full cloud session with its own branch and copy of the repo
• Project memory and instructions are shared, and new threads get them automatically
• Threads run in parallel and burn quota faster; the daily cap is 200 new threads
• The beta is open to some Pro and Max users only, limited to web and desktop, with no command-line support

This is a complete translation of the original article, rewritten sentence by sentence into natural English, with images taken from the original and supplemented by public screenshots from media coverage and project repositories. Original author: Michal Sutter, published September 17, 2026. Feature descriptions, quota caps and availability are translated as stated in the original, with nothing added or removed.