On October 10, 2026, Anthropic announced that everyone who previously applied for the Claude Code Projects beta with Pro and Max plans has received access. The announcement was published by the Claude developers' X account (@ClaudeDevs). The feature, launched on September 17, 2026 as a redesign of Projects, turns work in Claude Code into a 'coordinator' conversation: it automatically distributes tasks across parallel cloud threads that write code in separate git branches, run tests, and open pull requests. Access remains in beta, but it is now open to all applicants with paid plans.

image
image

What happened

On October 10, 2026, Anthropic announced via the Claude developers' X account (@ClaudeDevs) that the application review period has ended and the Claude Code Projects beta is open to everyone who applied with Pro and Max plans. The feature itself was launched on September 17, 2026 as a redesign of Projects: instead of the familiar folder of files and one chat, a 'coordinator' conversation works, which automatically breaks a task into parallel threads. Each thread is a separate cloud Claude Code session on its own git branch: it writes code, runs tests, opens a pull request, and then reports back to the coordinator. Threads share common project memory — the MEMORY.md index — and instructions up to 16,000 characters. A limit applies to the project: no more than 200 new threads per day.

Context

The original Projects in Claude Code was a modest wrapper around the model: a folder of files and one chat. The September 17 launch turned it into an orchestrator, architecturally structured as a planner-worker scheme without announced changes to the models themselves — this is a system layer on top of existing Claude Code, not a new model. Anthropic tested such orchestration on a closed whitelist for about three weeks and then opened it at once to all applicants with Pro and Max, but execution is currently possible only in the Anthropic cloud: local execution of threads is announced as 'coming soon,' without dates. Against this backdrop, the official communication framework is built around the 'Projects redesigned: from folder to conversation' post in the Anthropic blog and the Claude Code Docs 'Let Claude coordinate ongoing work with Projects' documentation.

Why this matters for the industry

For the industry, the event changes the status of the technology: the 'coordinator plus a swarm of parallel cloud sessions with shared project memory' orchestration has moved out of a closed pilot and is undergoing mass testing under real load from thousands of paid subscribers — effectively an externally non-reproducible field experiment. Essentially, this is a shift from 'one agent in a chat' to managing task threads as a product category, and the benchmark here is set for OpenAI Codex, Cursor, and GitHub Copilot, whose response in the coming months is almost inevitable. The economics of parallelism becomes a separate design axis: the cost of agent work in a swarm grows approximately linearly with the number of threads, which turns rate limits from a plan detail into a constraint on the approach itself. For startups, this is simultaneously validation of demand for multi-agent tools and a new UX pattern — 'goal setter reviews PRs' — which can be reproduced on top of their own stack.

Why this matters for users

The practical step is to check the Projects sidebar on claude.ai/code or the Code tab in the desktop app: applicants with Pro and Max should now see access. The scenario is then simple: describe a measurable goal — for example, 'reduce p75 checkout latency' — and connect a GitHub repository, after which Claude will create threads and open a pull request, and work will continue in the cloud even when the computer is off. It is important to keep in mind that each thread consumes plan limits as a separate full session, so parallelism quickly eats up the daily quota. The beta status means that results should be verified: it is reasonable to start with non-critical tasks and review open PRs before merging.

What is still unknown / limitations

The sources contain no metrics: the percentage of merged pull requests, reject rate, time to green CI, or comparison with a single session are not disclosed, so statements about a 'full-fledged autonomous agent' and 'almost asynchronous team' are not yet supported by measurements. It is not described how the shared MEMORY.md memory index is updated between threads and how edit intersections between parallel branches are resolved, so the risks of merge conflicts, memory degradation, and coordinator errors in task decomposition remain unassessed. Finally, the delivery dates for local thread execution and access expansion to Team and Enterprise are not named: what is announced as 'coming soon' remains a promise without dates, and the feature itself is in beta, where behavior may change.

Sources

Author

Look at AI, editorial team