Anthropic rolled out a redesigned Projects experience in Claude Code on September 17, 2026 — and it's a different product from the "Claude Projects" that's existed in claude.ai chat since 2025. The new version is a coordinator: one conversation that takes a stream of work, splits it into threads, runs each thread as a full parallel Claude Code cloud session, and keeps working after you close your laptop. It's in public beta today for select Pro and Max subscribers, rolling out wider over the coming weeks. Here's what a project actually is, how Anthropic says it's organized, and when it beats spinning up cloud sessions yourself.
This is a companion read to explainx.ai's coverage of Claude Code's agent view, cross-session messaging, and subagents and multi-agent workflows — three earlier steps toward the same problem this release addresses directly: coordinating more than one Claude Code session at once without doing the coordinating yourself.
TL;DR
| Question | Answer |
|---|---|
| What is it? | One long-running conversation where Claude acts as a coordinator, starting a cloud-session thread for each piece of work and tracking every thread it started |
| What's a thread? | A separate Claude Code cloud session on its own branch, with its own context window, that reports back to the conversation when it finishes or needs you |
| Who has access? | Pro/Max subscribers who already use cloud sessions and don't have existing chat/Cowork projects — beta rollout started September 17, 2026 |
| What does it need? | Optionally, GitHub repositories with the Claude GitHub App installed, plus files or Google Drive folders; a project works with zero repositories too |
| Does it work offline? | Threads run in the cloud and keep going when your computer is off — but they can't yet reach local files, a local database, or your internal network. Local support is "coming very soon" per Anthropic |
| What's the catch? | Running several parallel cloud-session threads draws on your Pro/Max usage limits faster than a single session — Anthropic explicitly flags this in the docs |
What a project actually is
Think of it as briefing a chief of staff instead of an individual contributor. You describe what needs doing — one task, several unrelated tasks, or a running stream of bug reports and requests over days — and Claude in the conversation decides what becomes a thread, whether an existing thread should pick it up instead, or whether it's a quick question it can just answer in place.
Anthropic's own examples in its launch post make the shape concrete: configure a project with a goal to reduce your app's checkout p75 latency, then ask Claude to profile each endpoint, test optimizations, and open pull requests in parallel threads. Or connect your API, web, and mobile repositories and set a goal to retire a deprecated v1 endpoint — Claude creates one thread per repository to migrate callers, run tests, open PRs, and then tells you which ones need to merge first, in what order.
Each thread is a full Claude Code cloud session: its own sandbox, its own branch, its own context window. A thread that needs to break its own assignment into smaller pieces can still use subagents, loops, and workflows internally — projects don't replace that layer, they sit a level above it, coordinating whole sessions rather than sub-tasks inside one.
How it's organized
A project has three parts that carry context to every new thread:
- The conversation — Claude acting as coordinator, taking what you send and deciding where it goes. It reads what threads report back, not every step they take internally, and works from recent messages, recent threads, and project memory rather than a full history — so it keeps going indefinitely without you managing its context window.
- Threads — the workers. Each clones the project's repositories, starts with the project's instructions and shared memory already loaded, does its assigned piece of work on its own branch, and opens a pull request when the work calls for one.
- The Overview pane — where every thread is grouped by state: Ready for review (PR open), Waiting on you (needs a reply, approval, or failed), Working, Landing (PR approved or merging), Idle, and Resolved. A dot on the Overview button tells you something needs attention without opening every thread to check.
Shared memory is the part that compounds. Every thread adds to and draws from the same project memory, so a fact the project learns once — the release moved to Friday, why an export got dropped, who to check with before touching the billing service — reaches every later thread automatically, instead of you re-explaining it in each new session. A Library tab collects both the files you upload and whatever files threads produce, so new work can build on what earlier threads already did.
When a project beats a single cloud session
A project earns its coordination overhead when the work has a goal that outlasts one session and keeps producing tasks — not for a single fix that fits in one sitting. Anthropic's own framing and the beta docs point at a few clear shapes:
- One goal spanning many repositories — "bring every service up to the new lint config," where each repository gets its own thread and its own PR.
- An area you keep feeding — the bugs, stack traces, and review requests for one service, pasted in as they arrive, with a pitfall you correct once staying fixed for every later thread.
- A build or migration bigger than one session — threads split the work, decisions made early reach later threads through shared memory, and new bugs found mid-build go into the same conversation instead of a separate one.
- Non-code work — a folder of contracts or a support-ticket export you keep coming back to with new questions; upload the documents instead of adding a repository.
For a single task that fits in one sitting — "fix the flaky login test" — a plain cloud session is still the right tool. Projects are for streams, not one-offs.
How it differs from Claude Tag, routines, agent view, and subagents
Anthropic now has several ways to run more than one Claude Code session at once, and it's easy to conflate them. The distinction that matters is who's coordinating and where the work runs:
| Feature | Coordinator | Runs where | Shared with |
|---|---|---|---|
| Projects | Claude, in the project conversation | Cloud (each thread is a cloud session) | You alone |
| Claude Tag | Whoever tags @Claude in Slack | Cloud | Your whole team/channel |
| Routines | A schedule, no conversation | Cloud | Whoever the project or session is scoped to |
| Agent view | You, manually | Local machine | You alone |
| Subagents | The parent session | Inside one session's context | That session |
A project can create a routine for scheduled slices of its work, and a thread can still spawn subagents for its own sub-tasks — the layers stack rather than compete. Cross-session messaging, which shipped in August 2026 so separate Claude Code sessions could message each other directly, solved a narrower version of the same coordination problem projects now handle by default: instead of you relaying a finding from one session to another, the project conversation does that routing itself.
Who has access, and what's still missing
The beta opened September 17, 2026 to Pro and Max subscribers who already use cloud sessions in Claude Code and have no existing projects in claude.ai chat or Cowork. Anthropic says access expands to more Claude Code users on those plans "over the coming week," then to all of Claude and to Team/Enterprise plans after that. If you're on Pro or Max without access yet, Anthropic points to a waitlist.
Two limitations are worth knowing before you lean on this for real work:
- No local reach yet. Threads run in Anthropic's cloud and keep working while your laptop is closed, but they can't currently touch files, a local database, or anything behind your internal network — only GitHub repositories and files you explicitly upload. Anthropic says local support is "coming very soon," but as of this beta, work that needs your machine specifically still belongs in a local session or agent view.
- Usage draws faster. Each running thread is a full session, and a project can run several at once with no fixed cap — Anthropic's docs are explicit that a project uses your plan's five-hour and weekly limits faster than a single cloud session, and recommend running fewer threads at a time or dropping to a smaller model for work that doesn't need the largest one.
Honest limitations
- Beta, gated rollout — most Pro/Max users don't have this yet, and Team/Enterprise access hasn't shipped at all.
- One project per user — you can't share a project or its threads with another person; there's no organization-level view during the beta, which is a real gap for teams that want project-level project.
- Threads can't be moved — a thread belongs to the project that started it; you can't relocate it to another project or spin it out to stand alone.
- Local tools are out of reach — the biggest practical gap for anyone whose workflow depends on a local database, device emulator, or VPN-only API.
What this means for builders
If you're already running two or three Claude Code cloud sessions in parallel and manually copy-pasting context between them, or checking back on each one individually, this is aimed squarely at that workflow. The value isn't parallelism itself — subagents and agent view already give you parallelism — it's that Claude does the coordinating: deciding what becomes a thread, remembering what earlier threads learned, and surfacing exactly what needs your attention instead of making you check every session yourself.
Related on explainx.ai
- Claude Code agent view and the /goal command: managing fleets of agents
- Claude Code cross-session messaging: sessions can now talk to each other
- Claude Code subagents and multi-agent workflows
- Claude Code loops official guide: turn-based, /goal, /loop, /schedule
- Claude Code commands: complete slash command reference
- Claude Tag: how to use Anthropic's team AI in Slack
- How to set up a Claude Project for your role (the earlier chat/Cowork feature)
- Official docs: code.claude.com/docs/en/claude-projects
This post reflects Anthropic's Claude Code Projects beta as announced and documented on September 17-18, 2026. Access, limits, and local-support timelines are rolling out in stages and may change — check Anthropic's own docs for the current state before planning work around them.