Anthropic just turned Claude Code Projects into a one-chat AI team. You drop in tasks, a coordinator spawns parallel threads that each do real work on their own branch, and you steer the whole thing from your phone. It’s in beta for Pro, Max, and Ultra subscribers right now, and access is rolling out in waves.
The problem this actually solves
If you’ve ever run two or three Claude Code sessions side by side, you know the drill. You paste the same context into each one. You flip between terminals to see who’s stuck. Basically, you play project manager for software that’s supposed to manage itself.
Anthropic noticed this too. As their models got faster, people kept spinning up more parallel sessions, and the whole experience started to crack. Attention gets fragmented. Context gets copy-pasted five times. Something drops.
The redesigned Claude Code Projects feature is their answer: stop managing sessions yourself. Hand the whole stream of work to one conversation and let it delegate. (We covered an earlier band-aid for this, cross-session messaging, when it launched. Projects goes a step further, because you don’t shuttle context between sessions at all.)
How the coordinator and threads work
Think of the project chat as a chief of staff. You brief it once. Then it breaks the work into pieces, spins up a thread for each one, and collects the results.
Under the hood, every thread is a full Claude Code cloud session. It gets its own copy of the repo and its own branch. It writes code, runs your tests, and opens a pull request when the work calls for one. If two threads touch the same files, the overlap shows up as a merge conflict, same as it would with human teammates. Nothing gets silently overwritten, which is exactly the detail I wanted to hear.
The Overview pane is where you watch it all. Threads get grouped into “working”, “ready for review”, and “waiting on you”. Cloud threads keep running after you close your laptop. And yes, you can nudge them from your phone while you’re in line for coffee.
Here’s how the five ways to run parallel work in Claude Code stack up, straight from the official docs:
| Approach | Who coordinates | Best for |
|---|---|---|
| Subagents | You, inside one session | Side tasks that would flood your main chat |
| Agent view | You, across sessions | Independent tasks you dispatch and check later |
| Agent teams | A lead agent (experimental) | Splitting one project among synced workers |
| Projects | Claude, from one conversation | Ongoing work spanning days or weeks |
| Dynamic workflows | A script | Huge audits and migrations with cross-checks |
Projects is the only one where Claude itself starts and tracks the sessions for you. That’s the whole pitch.
What every thread gets (and what it skips)
Each thread starts with the project’s repositories, the instructions you wrote once, the CLAUDE.md files, your account connectors, and a shared memory that grows as work happens. That memory is the quiet MVP here. Claude remembers the release moved to Friday and who to ping before anyone touches the billing code. We once walked through building your own AI memory system because chatbots kept forgetting things between sessions. Projects bakes that in for repo work, so half of that setup is now handled for you.
There’s also a library that collects the files you add plus the artifacts Claude produces. New threads build on old work instead of starting cold.
What threads don’t get matters too. They skip your repos’ skills, .claude/settings.json, and .mcp.json. If a thread needs those, the docs explain how to wire them into the cloud environment. It’s fifteen minutes of setup, honestly a little annoying, but it keeps the cloud sessions clean.
Who gets access right now
The rollout is gradual, so don’t panic if you don’t see it yet.
- Pro, Max, and Ultra: public beta, rolling out in waves. It started with accounts that had used cloud sessions and had no old-style projects.
- Where to look: claude.ai/code, the desktop app’s Code tab, or under “More” in the sidebar. The mobile app works too.
- Not yet: Team and Enterprise plans, the terminal CLI, the VS Code extension, and JetBrains plugins.
- No luck? Join the waitlist and wait your turn.
One more thing: if you still use old-style projects inside Claude chat or Cowork, those keep working. The redesigned experience reaches them later.
Set up your first project in about ten minutes
Setting up Claude Code Projects is deliberately boring. That’s a compliment.
- Open Projects at claude.ai/code and create one. Name it after the outcome, not the tool. “Cut checkout latency” beats “backend stuff”.
- Connect the repo, or drop in documents (more on that below). Claude reads the structure and suggests work it can start right away.
- Write the project instructions once. Target branch, code style, what never to touch. Every thread inherits this, so you set context a single time instead of ten.
- Give it a first goal. Something with a clear finish line: “profile every endpoint and open PRs for the three slowest”. Vague goals produce vague threads.
- Check the Overview pane after an hour. Review PRs, answer anything flagged “waiting on you”, and adjust how often the coordinator checks in. It adapts to how you like updates.
That’s it. The coordinator handles the juggling from there.
The catches nobody puts in the announcement
Your usage limits burn faster. Every thread is a full Claude Code session, so five parallel workers eat five sessions’ worth of tokens. Anthropic says the coordinator itself adds under 10% overhead, but the workers are the real cost. Claude Code Projects shows per-project usage, so you can see the burn before it hurts.
Uncommitted work can vanish. Cloud sandboxes pause between turns and resume when the thread continues. If a resume fails, the thread continues from a fresh clone, and uncommitted changes are gone. On long tasks, tell Claude to commit and push as it goes. Cheap insurance.
Projects are strictly personal. One user, no sharing, no organization controls during beta, and threads can’t move between projects. Fine for freelancers. Annoying for teams, until Team and Enterprise support lands.
Local threads have an asterisk. You can run a thread on your own machine through Remote Control (you may need to update Claude Code first), which lets it reach your local database or that API behind your VPN. But a local thread only works while your computer is awake. Cloud threads don’t care.
Do you need to be a developer?
Less than you’d think. Repos are the headline use case, but threads also work on documents. Drop in research files, briefs, or notes, and threads read them and draft.
Say you run a small business. You could hand a project three documents (last quarter’s numbers, customer emails, your pricing page) and ask it to draft a pricing update, a FAQ rewrite, and an email to existing customers, all at once. Each draft comes back as a finished piece of work you review in one place.
If you’d rather stay on OpenAI’s side of the fence, we wrote a playbook for building a real app with Codex that actually shipped. Different tool, same idea: describe outcomes, let the agent grind.
My honest take? Claude Code Projects is the first Claude feature that feels built for the way people actually work now: a pile of small related tasks, not one giant prompt. The beta quirks are real, but the direction is obvious.
What to do right now
Check whether Projects showed up in your sidebar at claude.ai/code. If it’s there, start small: one project, one repo, two tasks. If it’s not, join the waitlist and try the other parallel approaches in the meantime.
Parallel AI workers used to mean tabs, terminals, and hope. Now it’s one chat that remembers everything. That’s a real shift, and it’ll only spread across the rest of Claude from here.