The most useful sentence in Anthropic's September 17 announcement of the rebuilt Claude Code Projects is a warning, not a feature. Because a single project can now spin up several full Claude Code sessions simultaneously, the company says it will reach usage limits faster than ordinary work does. That trade-off defines the release: a coordinator that plans and delegates on the developer's behalf costs real tokens, and the beta is essentially an argument that the coordination is worth paying for.
What it replaces was passive. The old projects concept was closer to a labeled drawer β chats, files, and instructions grouped in one place, with nothing crossing between them on its own. The new one is an active participant that scopes a request, hands pieces of it to workers, checks what comes back, and stitches the result together while the developer is offline.
Key takeaways
- Each Claude Code Project now runs worker threads under a single coordinator, and every thread is a full cloud session on its own branch and copy of the repository.
- Threads share one project memory plus a library of uploaded files and generated artifacts, so a decision made in one thread reaches work started later in another.
- Anthropic set no separate price β projects draw on existing Pro and Max allowances, and several concurrent threads consume them faster.
What the coordinator actually does
Picture two layers. Workers sit at the bottom, and each one is a complete Claude Code cloud session with its own branch and its own copy of the repository. Above them sits a coordinator agent that keeps the goal, assigns work, and decides whether a new instruction belongs in a thread that already exists or deserves a fresh one. Anthropic's own analogy is briefing a chief of staff rather than issuing a command.
Isolation by branch has an obvious consequence the company does not try to hide. If two threads edit overlapping code, nothing clever happens β the result is a merge conflict, handled exactly as it would be on any other pull request. Workers can also delegate downward, splitting an assignment across subagents, loops, and workflows so that a wide migration proceeds in pieces.
Two scenarios from the launch post show the intended shape of the work. One sets a goal of reducing an application's checkout p75 latency, then lets Claude profile endpoints, test optimizations, and open pull requests in parallel threads. The other connects API, web, and mobile repositories to retire a deprecated v1 endpoint, with a thread per repository migrating callers and running tests before the coordinator reports which pull requests must merge first.
Why shared memory is the harder half
Running agents in parallel is table stakes at this point. Keeping them consistent is not, and that is where the project memory comes in: every thread both contributes to it and reads from it. Anthropic illustrates the payoff with the kind of detail that normally lives in a colleague's head rather than a repository β a release date that moved to Friday, an export that was quietly dropped, a service nobody should touch without asking a particular person first.
Around that memory sits a library holding whatever the developer uploads alongside whatever Claude generates, plus per-project settings for the cloud environment, connectors, plugins, instructions, and model. Check-in frequency and update verbosity are configurable too. One wrinkle is nomenclature: as VentureBeat pointed out, Claude has carried a separate Projects workspace since 2024, and that older product still shares knowledge without sharing operational state.
The limits worth reading before you start
Cost control is handled through visibility rather than caps. Projects expose their own usage view, and the coordinator and the worker threads take independent model and effort settings β so a cheap planner can direct expensive workers, or the reverse.
The bigger constraint is location. Threads execute in the cloud only, which is what lets them keep running after a laptop closes, but also means they cannot reach files, tools, or services inside a company network. Anthropic says local execution is arriving "very soon." Until it does, teams whose builds depend on private registries or internal APIs are watching rather than adopting. Mobile steering partly compensates: a developer can redirect a thread from a phone without keeping a workstation awake.
How it lands against Codex, GitHub, and Jules
Every serious competitor already runs coding agents concurrently. OpenAI's Codex app is built as a multi-agent workspace, GitHub anchors isolated sessions to branches, issues, and CI, and Google's Jules carries memory at the repository level across asynchronous tasks. The differentiator Anthropic is betting on is placement β a conversational coordinator above the workers, with memory scoped to an initiative rather than a repo. It extends the pattern behind other recent releases that gave Claude structured handles on outside systems, such as its browser-use tool.
Two questions will settle how this lands. Does the coordinator route well enough to beat a developer who divides the work manually, and does local execution ship before regulated engineering teams stop paying attention? Anthropic has moved the interesting problem from what a coding agent can build to who manages a room full of them.
FAQ
Who can use Claude Code Projects right now?
The beta opened on September 17 to selected Claude Pro and Max subscribers who use cloud sessions in Claude Code and do not already have projects on the web or desktop. Anthropic said access would expand to more Claude Code users on those plans within about a week, with Team and Enterprise plans following later.
Does Claude Code Projects cost extra?
No. There is no standalone price, and projects draw on the subscriber's existing Claude usage. Because each thread is a full Claude Code session and several can run concurrently, Anthropic warns that a project can hit usage limits faster than ordinary sessions.
Can project threads reach code on my own machine?
Not yet. Worker threads currently run as cloud sessions only, so they cannot touch local files, local tools, or systems behind an internal network. Anthropic says local execution is coming "very soon."
