Claude Code Projects: how parallel AI workers now run themselves
Key takeaways
- Each project thread is an independent Claude Code cloud session with its own repository copy and Git branch
- The coordinator now posts status updates unprompted, so developers no longer need to watch the thread list
- Parallel workers draw from the existing Claude Code allowance, and several high-effort sessions can exhaust limits fast
- Threads that touched a local computer cannot be shared, and Team and Enterprise plans are not yet supported
How do you run several AI coding agents at once without spending your day checking on them? Anthropic's answer, in the latest update to Claude Code Projects, is to make the coordinator speak first. It now posts status updates and answers questions in the project chat without being asked, and the release adds read-only thread sharing, reminders and two branching commands.
What Claude Code Projects actually is
A project is a long-running conversation with a coordinator, a Claude instance whose job is to direct work rather than do it. The work happens in threads. Each thread is an independent Claude Code cloud session, with its own copy of the repository and its own Git branch.
The simplest analogy is a site manager and a set of contractors. The manager holds the brief, hands out jobs and collects reports. The contractors work in separate rooms so they do not trip over each other. Anthropic describes briefing the coordinator the way you would brief a chief of staff, after which it routes work to new or existing threads.
When a project starts, the user picks a goal and a repository or other context, and Claude suggests work it can begin straight away. The cloud environment, connectors, plugins, instructions and model are all configurable.
How it works
The coordinator directs the project from thread reports. Because every thread has its own branch, overlapping changes surface as standard Git merge conflicts when branches are combined, just like any other pull request.
Anthropic gives two worked examples. In the first, a project has the goal of reducing checkout p75 latency, meaning the response time at or below which 75% of requests finish. Claude profiles each endpoint, tests optimisations and opens pull requests in parallel threads. In the second, a team connects its API, web and mobile repositories to retire a deprecated v1 endpoint. Claude creates one thread per repository to migrate callers, run tests and open pull requests, then reports which should merge first.
The new release removes the monitoring chore. Previously, developers had to scan the thread list to find sessions waiting for input. Now pending decisions, completed tasks and blocked work arrive in the main conversation. Other changes:
- Reminders: a thread can hide itself and resurface in the project chat at a scheduled time.
- /btw: opens a side thread for a tangent.
- /fork: branches a new thread from the current one. Both commands already existed in regular Claude Code sessions.
- Threads panel: a row of recent, updated and pinned threads sits at the top, and threads can show thumbnails of generated documents and diagrams.
- Sharing: a thread can be shared via a read-only link to anyone signed in to Claude, without exposing the rest of the project.
Anthropic's own materials add that every thread now draws on a shared project memory, and that a library collects files and artifacts. Each thread can also split its work further using subagents, loops and workflows.
Who uses it and why
The suited work is anything that divides into independent tasks and reconciles through branches and pull requests. Anthropic lists performance work, cross-repository migrations and batch changes such as refactors, dependency upgrades, test repairs and bug triage.
For a UK software house or in-house team, the migration case is the clearest fit: one deprecated interface, three codebases, and a merge order to work out. The value is less in speed per task than in not having to hold the whole picture in one's head.
What the limitations are
Cost is the main one. Projects carry no separate fee, but every worker consumes the subscriber's existing Claude Code allowance. New projects default to Opus for both roles, with low effort for the coordinator and high effort for workers. Several high-effort sessions running at once can exhaust plan limits faster than a single session. Accounts are capped at 200 new threads per day across all projects. A thread that hits a usage limit pauses and resumes automatically when capacity returns, and project usage appears separately in the usage controls.
Other constraints:
- Projects are cloud-only and limited to Claude Code. Anthropic says running threads on your own machine, alongside local tools and code and behind your network, is coming "very soon".
- Tasks with strict dependencies still need a developer to sequence pull requests and resolve conflicts.
- Small changes that fit one branch and one context window generally cost less in a regular Claude Code session.
- Threads that accessed a local computer cannot be shared, which stops transcripts exposing local file contents, shell output or other machine data.
Availability is narrow. The public beta began on 17 September for Pro and Max subscribers who joined the waitlist; Team and Enterprise are excluded. Anthropic says updated projects across all of Claude and on Team and Enterprise plans come later. The sources give no UK-specific pricing or regional availability, so UK subscribers should check their own plan and waitlist status.
What this means in practice
Developers on Pro or Max who already use cloud sessions should treat projects as a way to batch well-bounded jobs, not a replacement for a single focused session. Set the coordinator to low effort and reserve high effort for workers doing real analysis, since usage is shared. Watch the daily thread cap and the separate usage view before launching a large migration.
For IT leads, the Team and Enterprise exclusion matters more than the features. Until those plans are covered, and until local execution arrives, projects are an individual-subscriber tool. Teams comparing agent costs may also want to read how Sonnet 5.5, GPT-6.1 Sol and Gemini 4 Argon compare on cost per task, because parallelism multiplies whatever the per-task price is. For a different take on agents embedded in team workflows, see DeepSeek's dsh-superagent.
The read-only sharing rule is a sensible default for anyone handling client code: a link cannot leak what the thread never exposed.
Where this is heading
Anthropic's stated next step is running threads on a developer's own machine, alongside local tools and code and behind the company network. Rollout to chat, Cowork and the Team and Enterprise plans is the other milestone to watch. Whether usage allowances stretch to match the parallelism will decide how widely this is used.
Frequently asked questions
What is a Claude Code project?
A project is a long-running conversation with a coordinator that assigns work to threads. Each thread is an independent Claude Code cloud session with its own repository copy and Git branch.
Who can use Claude Code Projects?
The public beta is rolling out to select Pro and Max subscribers who use cloud sessions in Claude Code, on web, desktop and the latest mobile app. Team and Enterprise plans are not supported yet.
Does Claude Code Projects cost extra?
No separate fee applies. Every worker consumes the subscriber's existing Claude Code allowance, and accounts are limited to 200 new threads per day across all projects.
Can I share a Claude Code thread with colleagues?
Yes, through a read-only link for anyone signed in to Claude. Threads that accessed a local computer cannot be shared.