Claude Projects: A Smart Redesign as Grok Bot Gains Voice

Claude Projects has been rebuilt from a folder into a conversation, and Anthropic shipped the change on 17 September 2026 as a beta inside Claude Code. On the same day, xAI turned on voice mode for Grok Bot, its always-on agent product. Two companies, one working day, one shared idea: the thing you talk to should be the manager of the work, not the worker.

The Anthropic announcement is the larger of the two. A project used to be a container: some files, some instructions, and a chat that used them. The redesigned Claude Projects is a single ongoing conversation in which Claude acts as a coordinator, splitting what you ask for into parallel cloud sessions called threads, watching what they return, and carrying decisions forward in a shared memory that every new thread inherits.

The xAI announcement is smaller in scope and louder in delivery. Grok Bot, which launched in public beta on 11 August 2026, can now be spoken to rather than typed at, powered by the same Grok Voice engine xAI has been selling to developers. The official Grok account confirmed it at 20:00 UTC on 17 September with three words and a link.

This article covers what changed in Claude Projects, how threads and shared memory actually work, the confusing naming collision with the older feature of almost the same name, what it does to your plan limits, what Grok Bot voice mode is and is not, and what both launches say about where agent products are heading. Figures come from Anthropic’s announcement and documentation, xAI’s own posts, and the trade coverage published on the day.

Table of contents

What Anthropic Actually Changed in Claude Projects
How Coordination Inside Claude Projects Actually Works
What Every Thread Inherits in Claude Projects
Claude Projects and the Naming Problem Nobody Fixed
What Claude Projects Costs You in Plan Limits
Who Can Actually Use Claude Projects Today
xAI Launches Grok Bot Voice Mode the Same Day
The Enterprise Push Behind Grok Bot Voice
Why Both Launches Happened in the Same Week
What to Do Before You Turn Claude Projects On
Limits, Unknowns and What Nobody Has Confirmed
References

What Anthropic Actually Changed in Claude Projects

The Claude Projects redesign replaces a storage metaphor with a management one, and that is the whole point of it.

From folder to conversation

Anthropic’s own framing is the clearest summary available: “from folder to conversation”. The old container grouped chats and files. The new Claude Projects is one long-running session where Claude scopes the request, delegates the work, coordinates parallel threads, reviews the outputs and assembles the result.

The chief-of-staff comparison

In its launch thread, Anthropic compares briefing a project to briefing a chief of staff. You hand over several things at once, in any order, and Claude decides whether each one belongs in a new thread or an existing one, answering quick questions in place rather than spawning work for them.

Threads do the work

Each thread is a full Claude Code cloud session running on its own branch and its own copy of the repository. Threads open pull requests, run tests, and keep going after you close your laptop. Anthropic’s phrase for the architecture is “threads do the work, Claude directs it”.

Steering from a phone

Because the work runs in the cloud rather than on your machine, you can monitor and redirect it from a phone. That turns mobile into a supervision surface, which is a meaningful change for anyone who has had to leave a laptop awake to keep a long agent session alive.

The worked examples

Anthropic gives two. Set a goal to reduce an application’s checkout p75 latency, and Claude opens parallel threads to profile endpoints, test optimisations and raise pull requests. Or connect API, web and mobile repositories, ask it to retire a deprecated v1 endpoint, and it runs a thread per repository and reports which pull request has to merge first.

Layer	What it is	What it sees
The project conversation	One coordinating session with Claude	What threads report back, not every step
A thread	A full Claude Code cloud session	Its own branch and repository copy
Subagents, loops, workflows	How a thread splits its own assignment	Whatever the thread delegates
Project memory	Shared state written and read by threads	Decisions from across the whole project
The Overview pane	Thread states, library, pull requests	Everything that needs a human

How Coordination Inside Claude Projects Actually Works

The Claude Projects coordinator is not a scheduler with a progress bar. It is a conversation that happens to hold a work queue, and it talks back.

Routing, not queueing

When you paste a bug report or a stack trace into Claude Projects, Claude either answers it in place or starts a thread for it — and if a thread is already working in that area, it passes the task there instead of creating another one.

Thread states you can read

The Overview pane groups threads into Ready for review, Waiting on you, Working, Landing, Idle and Resolved. That vocabulary matters more than it sounds: it is the difference between “six agents are running” and “two of them need you”.

Pull requests that watch themselves

A thread opens a pull request when the work calls for one, then keeps watching it with auto-fix enabled. It pushes fixes when continuous integration fails and replies when checks pass, which is why an idle thread can wake up and spend usage hours later.

Merge conflicts stay merge conflicts

Coordination does not repeal source control. If two threads touch the same code, Anthropic says the overlap surfaces as an ordinary git merge conflict, resolved the way any other pull request conflict is.

Work that is not code

The documentation is explicit that a project can hold documents rather than repositories. Upload a folder of contracts or a support-ticket export, and threads deliver each write-up as a file on the project’s Library tab.

What Every Thread Inherits in Claude Projects

Standing context is the quiet half of this release, and arguably the more consequential one: it is what makes Claude Projects different from running six sessions yourself.

Project instructions

Each project carries instructions of up to 16,000 characters, sent to every new thread and to the coordinator. A rule you state once — target this branch, never touch that service — reaches all of them without being repeated.

Project memory

Claude writes and reads a memory index as the work proceeds. Anthropic’s examples are deliberately mundane: the release moved to Friday, the export was dropped, check with a named person before touching the billing service.

Repository context

Every thread clones each of the project’s repositories and loads the CLAUDE.md files, skills and plugins from all of them. Permission rules and hooks behave differently: they apply only from the directory a thread starts in, so a single-repository project honours them and a multi-repository project does not.

Connectors and tools

Model Context Protocol tools reach threads through the connectors on your claude.ai account. The project conversation itself has no connectors, which means connector work has to be sent to a thread rather than done by the coordinator.

The library

Files you add and artefacts threads produce collect in one library, so later work can build on earlier output instead of being re-fed it. This is the part that makes Claude Projects feel like a workspace rather than a queue.

Claude Projects and the Naming Problem Nobody Fixed

There is a genuine ambiguity in this launch, and it is worth stating plainly rather than glossing over.

Two products, almost one name

Anthropic already had Claude Projects: the general workspace, introduced in June 2024, that groups chats, files, instructions and project knowledge. The thing launched on 17 September is the redesigned projects experience inside Claude Code, which trade coverage has variously called Claude Code Projects.

How Anthropic itself words it

The company’s blog post is titled “Projects redesigned: from folder to conversation” and describes “a new experience for Claude projects, now available in beta in Claude Code”. So the new thing is not a separate product line — it is the same feature name, rebuilt, starting in the developer surface.

What the difference amounts to

The older workspace helps Claude remember the context you gave it. The redesigned Claude Projects helps Claude remember and coordinate the work it is doing. Knowledge versus operational state is the distinction that matters when you are choosing which one to open.

Why the confusion is temporary

Anthropic says existing projects on Pro and Max keep working as they do today and will be upgraded as the rollout expands to chat and Cowork. The two-products-one-name period is a rollout artefact, not a permanent product split.

What to call it internally

If your team already says “projects” for the old workspace, pick a temporary qualifier now. Retrofitting a name after half the engineering organisation has adopted the new Claude Projects is harder than agreeing one in advance.

What Claude Projects Costs You in Plan Limits

Claude Projects has no standalone pricing, which is exactly why the usage arithmetic deserves attention.

No new price, faster burn

Claude Projects draws on the subscriber’s existing plan. Because every thread is a full Claude Code session and several run at once, Anthropic warns that a project reaches usage limits faster than a single session does, and calls out Pro plans specifically.

The one hard ceiling

There is no fixed cap on concurrent threads; Claude starts as many as the work calls for, and a limit you request is treated as a preference. The enforced ceiling is 200 new threads per day across your projects.

Default model settings

A new project runs Opus everywhere, with high effort for threads and low effort for the coordinator conversation. That is the most expensive default combination available, and it is adjustable in project settings.

Threads that wait rather than fail

A thread that hits your plan limit waits and resumes on its own when the limit resets. Work you left running therefore starts consuming your next usage window without telling you first.

Where to look

Each project exposes its own usage tab, plus separate model and effort controls for the coordinator and for threads. Checking it after the first real project is the cheapest habit to build.

Claude Projects: stated limits and defaults (Anthropic documentation, 17 September 2026)

New threads per day, all projects 200

Project instructions, characters 16,000

Concurrent threads no fixed cap

Standalone price none

Bars are scaled to the 200-thread ceiling; the 16,000-character figure is shown at 62% for legibility, not as a ratio of threads.

Who Can Actually Use Claude Projects Today

Access to Claude Projects is narrower than the headlines suggest, and the gating criteria are unusually specific.

The launch cohort

The beta opened to select Claude Pro and Max subscribers who already use cloud sessions in Claude Code and who do not have existing projects on the web or desktop. If that describes nobody on your team, nothing appears in the sidebar.

The expansion path

Anthropic said access would widen to more Claude Code users on those plans over the following week, with the redesigned experience reaching the rest of Claude, plus Team and Enterprise plans, after that. A waitlist exists for Pro and Max accounts without access.

The GitHub requirement

For code work, the documentation is strict: repositories on github.com rather than GitHub Enterprise Server, GitLab or Bitbucket, a connected account with push access, and the Claude GitHub App installed. A terminal-issued web setup token is not sufficient for project threads.

Cloud now, local later

Threads run in the cloud, and Anthropic states plainly that they cannot yet reach files or tools on your machine or services on an internal network. Local execution is described as coming “very soon”, with no date attached.

Who should wait

If your build depends on private package registries, internal services behind a VPN, device emulators or local databases, the current release cannot run it. That is a scoping constraint, not a bug, and it is the main reason to hold off.

xAI Launches Grok Bot Voice Mode the Same Day

The second half of the day belongs to xAI, and it is a smaller change than Claude Projects with a wider audience.

What was announced

On 17 September, Grok Bot gained voice mode: users speak to their bots instead of typing. The Grok account’s confirmation was three words — “Powered by Grok Voice” — posted at 20:00 UTC, roughly 45 minutes after the first user report of the feature appearing.

What Grok Bot is

Grok Bot launched in public beta on 11 August 2026 as a team of always-on agents that work in the background. Bots can run in parallel, one can manage others, they message each other, and they can sit in a group chat together. xAI’s own description is that they “pass work, assign ownership, and only pull you in for judgment calls”.

The voice engine underneath

Grok Voice is xAI’s real-time speech-to-speech layer, which collapses transcription, reasoning and synthesis into one model rather than chaining three. The current generation, Think Fast 2.0, was announced on 29 July 2026 and cut time to first audio to roughly 0.70 seconds.

The developer release a day earlier

On 16 September, one day before the consumer feature, xAI made Grok Voice available on the fal platform: bidirectional WebSocket streaming, more than 25 languages, 26 voice options, audio cloning, and tool calling mid-conversation, priced at $0.00083 per second — about $3 per hour of continuous audio.

Platforms and caveats

Grok Bot runs on desktop and iOS, with Android support added in early September 2026. xAI has not published an availability matrix for voice mode specifically, so which cohort sees it first is unconfirmed.

Dimension	Claude Projects (Anthropic)	Grok Bot voice (xAI)
Announced	17 September 2026, blog and docs	17 September 2026, posts on X
Primary surface	Claude Code on web and desktop	Grok app, desktop and iOS
Unit of work	A thread, one cloud session per task	A bot, persistent and always on
Interface change	Folder becomes a conversation	Typing becomes speaking
Shared state	Project memory and library	Bots message each other directly
Documentation	Full docs page at launch	No published feature page

The Enterprise Push Behind Grok Bot Voice

Voice mode did not arrive in isolation. It landed on the last day of a three-day xAI event with an unusual format.

The 72-hour company

At The Howard in San Francisco, three xAI staff — Lauren Tan, Matt Palmer and Roshan Sadanani — spent 72 hours building a functioning company whose workforce was Grok Bot. The demonstration closed on 17 September, the same day voice mode appeared.

Three enterprise products

The event shipped a Voice Agent API at $0.08 per minute for speech-to-speech, the no-code Voice Agent Builder that has been in beta since 1 July 2026, and an Agent Tools API for wiring agents into proprietary software.

The pricing signal

At $0.08 per minute, xAI is pricing below the established speech vendors it names as competition. Undercutting on the infrastructure layer while giving the same engine away inside a consumer feature is a deliberate commoditisation play, and it is worth reading as one.

A procurement bot with a number attached

xAI also pointed to Haggle Bot, launched on 3 September 2026, which monitors vendor spend and contracts. The company says it has generated more than $100,000 in direct savings internally — a real figure, from a single interested party, in a company with unusual tolerance for automation.

The architecture claim

Each agent runs in its own virtual machine under what xAI calls a VM-per-agent architecture, backed by a stated $50 million investment in agent security. Access, network and audit controls, action recording and telemetry export are positioned as the enterprise-only layer.

Why Both Launches Happened in the Same Week

Two very different companies moved the same interface boundary within hours of each other, and the reason is structural rather than coincidental.

The bottleneck moved

Running one agent is solved. Running eight and reconciling what they did is not. Both releases attack the coordination cost rather than the generation cost, which is where the friction has actually been for the past year.

Everyone is converging on the same primitives

OpenAI’s Codex app is the closest comparison to Claude Projects: parallel agents, isolated changes, supervision of longer work. GitHub anchors the same idea to branches, issues and continuous integration. Google’s Jules keeps repository-level memory across asynchronous tasks.

The differentiator is where control sits

Anthropic’s distinctive choice is an explicit coordinator above the workers, with one conversation as the top-level interface and shared memory spanning threads. xAI’s is a persistent agent you can now simply talk to. Same problem, opposite ends of the same spectrum.

Voice is the other half of the same idea

If an agent runs for hours without you, the cheapest way to intervene is speech. Grok Bot voice mode and steering Claude Projects from a phone are the same admission: supervision has to be possible when you are not at a keyboard.

What both launches assume about instructions

Both assume that natural language processing is now reliable enough to serve as the control surface for multi-step work, rather than a convenient front end on a form. That assumption, not raw model quality, is what Claude Projects and Grok Bot are really testing, and it is the thing to watch over the next quarter.

What to Do Before You Turn Claude Projects On

A short, practical sequence for teams evaluating Claude Projects in the next fortnight.

Scope one stream of work, not a backlog

A project suits a goal that outlasts a session and keeps producing tasks: one service’s bugs, one migration, one latency target. Pointing Claude Projects at an unbounded backlog is how you meet the 200-thread ceiling without learning anything.

Write the instructions before the first thread

You get 16,000 characters that reach every thread. Branch policy, review expectations, the services nobody touches without asking — these belong there on day one, not after the third surprising pull request.

Turn the model defaults down first

Opus at high effort on every thread is the default and the most expensive setting. Start lower, watch the usage tab, and raise it where the work genuinely needs it.

Check the GitHub app, not just the token

The single most likely setup failure is a connected GitHub account without the Claude GitHub App installed on the repository. Confirm that before wondering why threads cannot push.

Decide who reviews the pull requests

Parallel threads produce parallel pull requests. If your review capacity is one person on Friday afternoon, Claude Projects will find that out faster than any process document ever has.

Limits, Unknowns and What Nobody Has Confirmed

Both stories have edges that the coverage tends to smooth over. These are the ones worth keeping in mind before you rely on Claude Projects for anything load-bearing.

Cloud-only is a real constraint

Until local execution ships, threads cannot see your machine, your VPN or your internal registries. “Very soon” is the only timing Anthropic has offered.

Permission rules behave inconsistently

Hooks and permission rules apply only from the directory a thread starts in. In a multi-repository project they do not apply across the other repositories, which is a sharp edge worth testing deliberately.

Grok Bot voice has no documentation

There is no xAI feature page, no availability matrix and no stated model version for the consumer voice mode. The Think Fast 2.0 attribution is inference from xAI’s developer releases, not a confirmed statement about the bot feature.

Vendor figures are vendor figures

The $100,000 Haggle Bot saving, the latency numbers and the transcription comparisons all come from xAI. Anthropic’s thread-state model and limits come from Anthropic. Neither set has been independently verified.

The usage risk is asymmetric

An idle thread that wakes on a failed continuous integration run spends your quota without a prompt from you. On a Pro plan, that is the failure mode most likely to surprise a first-time user of Claude Projects.