Claude Code Projects Is a Public Beta Documented by Anthropic, and Subagents Already Run in Their Own Context Window

The feature this video calls "secret" has a public documentation page at code.claude.com, a claude.com announcement dated 2026-09-17, and an official @ClaudeDevs post from 2026-10-09 saying every Pro and Max user on the waitlist had been let in, and the video's own B-roll shows two of those public pages on screen. The artifact is a 74.07-second vertical clip, 1080x1920 at 9:16, in an MP4 container (ffprobe reports mov,mp4,m4a,3gp,3g2,mj2), video in HEVC Main as hvc1 at 343,839 bps and 60 fps across 4,443 frames, audio in HE-AACv2 at 44,100 Hz stereo and 48,120 bps, 400,777 bps overall, 3,710,749 bytes on disk. TikTok's metadata puts the post at 2026-10-17 01:08:32 UTC from uploader chase_ai_, channel nickname "Chase AI", with the audio credited as "original sound - Chase AI" and a description identical to the title. At capture on 2026-10-19 the counters read 8,253 views, 299 likes, 15 comments, 49 reposts, and 245 saves. I read all 37 extracted frames and the 252 transcript words in the 23 segments that fall inside the 74.14-second audio; Whisper appended five segments of lone periods starting at 86.74 seconds, past the end of the audio, and I discarded them. Whisper also heard "Cloud Code" and "Cloud Projects" throughout; the speaker says "Claude", and the on-screen caption in frame 32 reads "so Claude Projects". Every frame is a vertical split: a screen recording in the upper half, a one-to-three-word caption in white and red bold type across the middle, and in the lower half a bearded presenter with short brown hair in a cream henley, a black microphone in front of him, a door at left, and a framed orange-and-blue city skyline print behind him. Frames 1 to 3 show a Claude Code desktop window open on a repository named checkout-service, with a chat pane on the left and a "Threads" panel on the right reading "Welcome back, Lydia." and "Nothing is waiting on you." The panel lists three threads under "Working 3" ("Fix /checkout p99 regression", "Draft release notes", "Bump stripe-node to v15") and one under "Idle 1" ("Upgrade pg driver to 8.13", "PR #4790 merged, CI green on main", "2h"). The user's 9:12 AM message reads "p99 on /checkout doubled after deploy #4812, find it and fix it. Also draft Friday's release notes, and bump stripe-node to v15." and Claude replies "On it. Three threads; I'll flag anything that needs you." with the tag "Sent to 3 threads". The model selector at the bottom reads "Fable 5.1" and "High". Frames 4 to 6 show the code.claude.com docs page "Let Claude coordinate ongoing work with Projects" under the breadcrumb "PLATFORMS AND INTEGRATIONS", with a blue callout beginning "Projects are in public beta on Pro and Max plans". Frames 7 to 9 show a claude.com page headed "Projects are becoming a conversation with Claude" with the nav items Product, Developers, Enterprise, Resources, and Pricing. Frames 10 to 12 show the waitlist form: First name, Last name, "Claude account email" with the hint "If you're an existing user, please enter your account email.", "Company or organization name", and "What do you like to use Projects for?" with checkboxes for Coding, Knowledge Work, and Daily Tasks. Frames 13 and 14 show the desktop app sidebar: Search, New, Projects with a "Beta" badge, Artifacts, Scheduled, Customize, Pinned, The Vault, and a truncated project row beginning "chase-ai-ca". Frames 15 to 21 and 28 to 31 show a hand-drawn panel in a whiteboard app headed "WITH PROJECTS / Claude is the project manager": a stick figure labeled "YOU" says "build these 4 things. I'm going to bed." under the note "one message", an arrow leads to "PROJECT CHAT / the manager ('coordinator') / splits the work, routes follow-ups, reads every report", which fans out inside a cloud labeled "ANTHROPIC'S CLOUD / keeps running when you leave" to four terminals ("build signup page", "fix login bug", "write tests", "update docs"), each with a green "PR opened" tag, all drawing from a yellow "SHARED BINDER (every thread starts with it) / project instructions · memory · CLAUDE.md · skills · connectors", with the footnote "each thread = a full Claude Code session · own branch · fixes its own CI" and the bottom tags "lid closed = still shipping" and "check in from your phone". Frames 22 to 27 show the companion panel "TODAY / you're the project manager": a stick figure thinking "wait... which tab was the bug fix?", four terminals each tagged "waiting on you", a box reading "your job, every time: 1. write 4 separate briefs 2. paste the same rules 4x 3. copy context between tabs 4. check every single result 5. open + merge every PR 6. keep the laptop open", and the line "close the laptop = everything stops". Frames 32 to 37 show a third panel, headed in part "actually different?" and "how long it lives", with "A PROJECT" over "MANAGER" over three workers each labeled "own context" with a progress bar and "own PR", the note "each worker is a SEPARATE full session", and the line "Claude does. You drop work in, the manager hands it out." The presenter draws circles and arrows over these panels with a pen tool as he talks.

The feature is documented in public, and the video shows the pages that document it

The narration opens with "Claude Code has a secret feature hidden behind a waitlist." I am correcting that framing. The page in frames 4 to 6 is Let Claude coordinate ongoing work with Projects on Anthropic's Claude Code documentation site, which I fetched on 2026-10-19. Its callout reads, in full: "Projects are in public beta on Pro and Max plans and rolling out gradually, starting with accounts that have used cloud sessions and don't have existing projects in claude.ai chat or Cowork. They aren't available on Team or Enterprise plans yet. If Projects doesn't appear in the sidebar at claude.ai/code or in the Code tab of the desktop app, the rollout hasn't reached your account, and you can join the waitlist. Run agents in parallel lists what you can use in the meantime."

The page in frames 7 to 9 is claude.com/form/projects, whose static HTML carries the heading "Projects are becoming a conversation with Claude" and the sentence "We are rolling out Projects in stages. Join the waitlist to get earlier access." Anthropic also published Projects redesigned: from folder to conversation on 2026-09-17, which states: "Starting today, updated projects are available in beta to select Claude Pro and Max subscribers who use cloud sessions in Claude Code." A public beta with a documentation page, a marketing page, a dated announcement, and a sidebar entry in the desktop app (frame 13, "Projects" with a "Beta" badge, which the desktop docs describe as "Select Projects in the sidebar to create a project") is a gated rollout. Gated and secret are different things, and the video's own footage shows the public pages.

One naming point. The video says "Claude Projects". Projects in claude.ai chat have existed as a separate feature, which the support article describes as "self-contained workspaces with their own chat histories and knowledge bases." The feature in this video is the Claude Code redesign of that idea, which the docs call projects and the Oct 9 @ClaudeDevs post calls "Claude Code Projects". The claude.com page bridges the two: "It's new in Claude Code and an upgrade for the projects you already have in chat and Cowork."

The waitlist is real, the 24-hour approval is the uploader's experience, and the plan requirement goes unmentioned

The URL the narration spells out, "claude.com slash form slash projects", resolves to the page described above, and the docs callout links to the same address as "join the waitlist". The form fields in frames 10 to 12 match a HubSpot-embedded form; the static HTML I fetched contains HubSpot markers and the "Coding" option, and the "Knowledge Work" and "Daily Tasks" labels are visible only in the frames, since the form renders client-side.

The narration promises "You get off the waitlist in less than 24 hours." No primary source states an approval time. The docs callout says "rolling out gradually", the Sept 17 article says "select", and the support article says "select Pro and Max subscribers". The closest primary statement is the @ClaudeDevs thread of 2026-10-09, which I could only reach through a Thread Reader mirror because x.com returned HTTP 402 to my fetches: "We just let in every Pro and Max user from the Claude Code Projects waitlist!" followed by "Projects are in beta. If you haven't signed up yet, join the waitlist. We're letting in more Pro and Max users as capacity allows." The video went up on 2026-10-17, eight days after that post, so a fast turnaround is plausible for the uploader's account. I am labeling the under-24-hours figure unverified, since it is one person's experience and Anthropic conditions admission on capacity.

The video never says who can apply. The docs are specific: Pro and Max plans only, and "They aren't available on Team or Enterprise plans yet." The cloud sessions page adds a prerequisite for code work: "threads in a project need the Claude GitHub App installed on each repository they clone, whichever method you connected with." A viewer on a Team plan, or with repositories on GitLab, will fill out the form and wait for a feature the docs say they cannot use yet.

"Runs 24/7 with your computer off" holds for cloud threads, within plan limits and a 200-thread daily cap

The narration says "It runs 24-7. It doesn't matter if your computer is off." The docs support the core claim: "Cloud threads keep going after you close your laptop, while a thread on your computer runs only while that computer is awake." The cloud sessions page says the same of any cloud session: "The session keeps running after you close your laptop, and you can check on it or steer it from any device."

Three constraints shape what "24/7" means in practice, and the video mentions none of them. First, plan limits. The docs state: "A project uses your limits faster than a single session does, and on a Pro plan in particular you should expect to reach your limit sooner on days you run one." A thread that hits the plan's five-hour or weekly limit "keeps retrying on its own and continues when the limit resets", so the work pauses rather than stops, and it resumes into your next usage window without asking. Second, a hard cap: "The enforced limit is 200 new threads per day across your projects." Third, idle expiry: cloud sessions "stop after a period of inactivity and the session's VM is reclaimed", and reopening restores the conversation history but not "background work that was still running when the VM was reclaimed."

The narration's closing line, "you want to talk with one agent instead of 20", has no fixed number behind it. The docs say "There's no fixed number; Claude starts as many as the work calls for, and a limit you ask for is a preference rather than a cap", and they recommend that a new user "Ask Claude to propose threads before starting them and to run a few at a time, and drop those limits once a few threads come back the way you want."

Threads share memory and instructions, and each one has its own context window

The narration says "all these agents share memory and context", then 20 seconds later says each session has "their own context window". Both halves have a precise reading in the docs, and they do not conflict once the terms are pinned down.

What threads share: the page subtitle says threads are "parallel cloud sessions that share repositories, instructions, and memory." Project memory is a set of files, and "Every cloud thread reads the index file MEMORY.md when it starts and opens the other files when it needs them." Project instructions are "Text sent to each new thread and to Claude in the project conversation, up to 16,000 characters." Every cloud thread also starts with the project's repositories, the CLAUDE.md and skills in each of them, and the connectors on the user's claude.ai account. This is exactly the "SHARED BINDER" in the diagram at frames 15 to 21, which lists "project instructions · memory · CLAUDE.md · skills · connectors". That drawing is accurate.

What threads do not share: a context window. The docs define a thread as "a separate session with its own context window that does one piece of work and reports back to the conversation when it finishes." The coordinator "sees what threads report back, not every step they take." So "share context" is true only in the sense of shared standing context (files, memory, instructions), and false in the sense of a shared conversation. The docs also note one exception the video does not cover: a thread run on your own computer through Remote Control "Starts with the project's instructions, but not with its memory files loaded."

The subagent comparison picks the wrong differentiator

The narration says "this is different than sub-agents, since all of these sessions and projects that are coordinated in the cloud have their own memory, their own context window, and they are persistent." Of those three, the context window is shared ground, so I am correcting it as a differentiator. The subagents docs state: "Each subagent runs in its own context window with a custom system prompt, specific tool access, and independent permissions."

The docs draw the line elsewhere. From the Projects page: "a subagent runs inside one session, does a side task in its own context window, and returns a summary to that session. A project's threads are whole sessions that Claude starts and that report back to the project conversation, and a thread can still use subagents for its own side tasks." The Run agents in parallel comparison table describes subagents as "Delegated workers inside one session that do a side task in their own context and return a summary", and projects as "One ongoing conversation at claude.ai/code or in the desktop app. Claude starts parallel sessions called threads, in the cloud or, when you ask, on your computer through Remote Control, gives each one the project's instructions, and shows you which ones need you."

The persistence point survives. The subagents docs say "Each subagent invocation creates a new instance rather than continuing an earlier one", while a thread is a session you can open, steer, and send follow-ups to. The memory point survives with a caveat: a non-fork subagent does not load the main conversation's auto memory by default, while every cloud thread reads project memory at start. The third diagram at frames 32 to 37 ("each worker is a SEPARATE full session", "own PR") states the real difference better than the narration does.

The hand-drawn diagram matches the docs, and the UI label matches Anthropic's current model list

The "WITH PROJECTS" panel's footnote, "each thread = a full Claude Code session · own branch · fixes its own CI", lines up with the docs: "A cloud thread works on its own branch and opens a pull request when the work calls for one." The "PROJECT CHAT" box's description of the manager as the one who "splits the work, routes follow-ups, reads every report" matches "It takes what you send, decides what becomes a thread, and keeps track of every thread it started." The "TODAY" panel's six-item list of manual chores paraphrases the docs' own framing: "Without a project, running several sessions means doing the coordinating yourself: you decide what each one works on, repeat the same background at the start of each, and check back to see which finished or needs an answer."

One thing the diagram's cloud, labeled "ANTHROPIC'S CLOUD / keeps running when you leave", leaves out is that threads can now run locally. The Sept 17 article said "Threads run in the cloud today; running on your machine alongside your local tools and code and behind your network is coming very soon." By the time of my fetch the docs had a full section, "Run a thread on your own computer", requiring Claude Code v2.1.280 or later and Remote Control, with the thread pausing "while that computer is asleep". The video, posted a month after the announcement, describes the Sept 17 state.

The model selector in frames 1 to 3 reads "Fable 5.1" and "High". Anthropic's models overview, fetched 2026-10-19, lists Claude Fable 5.1 with API ID claude-fable-5-1, a 1M-token context window, and a default effort of high, so the label is a real model at its default effort level. The docs say these settings apply per project: "Coordinator model and Coordinator effort apply to Claude in the project conversation." The greeting "Welcome back, Lydia." identifies a demo account, and I did not find a primary source for whose screenshot it is.

Key Takeaways

Correction: Projects in Claude Code is a public beta with a documentation page, a claude.com announcement dated 2026-09-17, a desktop sidebar entry labeled Beta, and an official Oct 9 post about waitlist admissions; the video's own B-roll shows the public pages it calls secret.
Verified: The waitlist lives at claude.com/form/projects, access is limited to Pro and Max plans, Team and Enterprise are excluded for now, and threads need the Claude GitHub App on each repository they clone.
Unverified: "Off the waitlist in less than 24 hours" is the uploader's experience; @ClaudeDevs said on Oct 9 that every waitlisted Pro and Max user had been admitted and that further admissions depend on capacity.
Clarification: Cloud threads keep running after you close your laptop, subject to plan usage limits that pause threads until the window resets, an enforced cap of 200 new threads per day, and VM reclamation after inactivity.
Clarification: Threads share repositories, instructions, and project memory, and each thread has its own context window; the coordinator sees thread reports rather than every step.
Correction: Subagents also run in their own context window, so that is no distinction; the docs separate the two by scope (a side task inside one session versus a whole session Claude starts and tracks) and by persistence.

Resources

Let Claude coordinate ongoing work with Projects establishes the public beta callout, the Pro and Max requirement, the thread and coordinator definitions, the shared memory and instruction mechanics, the 200-thread daily cap, the local-thread section, and the subagent comparison.
Run agents in parallel establishes the five-way comparison table that places projects beside subagents, agent view, agent teams, and dynamic workflows.
Create custom subagents establishes that each subagent runs in its own context window, that each invocation is a new instance, and that auto memory is not loaded by default.
Use Claude Code in the cloud establishes that cloud sessions keep running after you close your laptop, that they expire after inactivity, and that project threads need the Claude GitHub App on each repository.
Desktop application establishes the Projects item in the sidebar seen in frames 13 and 14.
claude.com/form/projects establishes the waitlist heading, the "rolling out Projects in stages" sentence, and the HubSpot form; no approval timeframe appears on it.
Projects redesigned: from folder to conversation establishes the 2026-09-17 announcement, the "select Claude Pro and Max subscribers" wording, and the statement that local threads were "coming very soon" at launch.
What are projects? establishes the older chat-based Projects feature and the "select Pro and Max subscribers" wording for the new version.
Thread Reader copy of the @ClaudeDevs thread of 2026-10-09 establishes the "We just let in every Pro and Max user from the Claude Code Projects waitlist" statement and the "as capacity allows" caveat; the original at x.com/ClaudeDevs/status/2108621476538781878 returned HTTP 402 to my fetch.
Models overview establishes that Claude Fable 5.1 (claude-fable-5-1) is a current model with a default effort of high, matching the UI label in frames 1 to 3.