Anthropic ran Code with Claude Tokyo on June 10, 2026, the third stop of its 2026 developer conference after San Francisco and London. Day 1 was a coordinated launch moment for Claude Code, not a single product reveal. The keynote and the "What's new in Claude Code" session pulled together a long list of capabilities into one narrative: Claude Code is no longer a terminal CLI you babysit. It is a desktop application, a cloud-triggered automation surface, a voice-driven assistant, and a fleet of parallel sub-agents you can run from your phone. Anthropic dropped Claude Fable 5 the day before so Tokyo could demo every new feature on the new model.
This is a builder-focused recap of what Anthropic showcased at Code with Claude Tokyo Day 1, what each feature actually does, and how the new Claude Code platform shape changes how agencies and SaaS teams ship.
Quick Answer
- Where: Code with Claude Tokyo, June 10, 2026. Day 2 Extended (June 11) targets founders and indie developers with builder stories.
- Headline: Claude Code is becoming a platform. Anthropic packaged Desktop, Auto Mode, Worktrees, Auto memory, Voice Dictation, Remote Control, Routines, Dynamic Workflows, Agent Teams, Managed Agents Vault, and Security Scanning into a single product surface.
Read the production playbook for shipping Anthropic Managed Agents in 2026 for the schedule + sandbox + MCP-tunnel + outcome-grader pattern.
- Underlying model: Claude Fable 5 (general availability June 9, 2026) is the default backbone for all Claude Code surfaces. See our Claude Fable 5 deep dive for the model details.
- For agencies and SaaS teams: Routines plus Dynamic Workflows turn Claude Code into an unattended worker that reacts to GitHub events, cron schedules, and HTTP triggers. That is the cleanest path so far to running Claude Code at fleet scale.
- For Totalum users: every feature here makes the agent loop richer. If you connect Claude Code to Totalum via MCP, the agent can now build, ship, and maintain real Next.js apps from a desktop, a phone, or a scheduled webhook.
Code with Claude Tokyo, in context
Code with Claude is Anthropic's traveling developer conference. The 2026 edition ran in San Francisco on May 6, London on May 19, and Tokyo on June 10. Each stop has three parallel tracks: Research, Claude Platform, and Claude Code. Tokyo Day 1 ran the same triple-track format with 28 sessions, headlined by a 60-minute opening keynote featuring Angela Jiang (Head of Product, Claude Platform), Cat Wu (Head of Product, Claude Code), Dianne Penn (Head of Product, Research), and Katelyn Lesse (Head of Engineering, Claude Platform).
Tokyo Day 2, branded as Code with Claude Extended, targets independent developers and early-stage founders with stories like "Manager 11: How I finally shipped my dream football management game with Claude" and "Rewriting Bun in Rust." Day 2 is a community day, not a product launch day, so the news in this recap comes from Day 1.
The key framing Anthropic used at Tokyo: Claude Code is no longer just the CLI in your terminal. It is a Code tab inside a desktop application, a routine running in the cloud, an agent your phone can wake up, and a fleet of sub-agents working a single ticket in parallel.
The 11 Claude Code updates from Tokyo Day 1
1. Claude Code Desktop is the new default surface
Anthropic showcased the Claude Desktop app as the new home for Claude Code. It now has three tabs: Chat for conversations, Cowork for long-running agentic work, and Code for software development. Each session in the Code tab has its own chat history, project folder, and code changes, and runs in parallel with other sessions.
This matters because the previous mental model was "open a terminal, run claude, work on one task." The new model is "open the Claude app, run six sessions side by side, each on its own Git worktree, each with its own pane layout, each capable of being scheduled or remote-controlled."
The desktop app ships for macOS (Intel and Apple Silicon), Windows x64 (with a separate ARM64 build), and Linux users still get the CLI.
2. Auto Mode arrives without a terminal in front of you
Auto Mode lets Claude Code continue through tool calls without an approval prompt for every action. A classifier judges each tool invocation, auto-continues the safe ones, and stops to ask only for risky operations. It was first announced earlier in 2026 for Max, Team, and Enterprise; Tokyo positioned it as a default for the Desktop app.
The practical effect: a 90-minute migration job stops being "click Yes 200 times" and starts being "let Claude finish, then review the diff." If you have read our Claude Code subagents writeup, Auto Mode is the gate that finally lets parallel subagents run unattended for hours.
3. Worktrees end the branch-switching tax
Worktrees give each Claude Code session its own isolated Git worktree, so two sessions on the same repo never stomp each other. The Desktop app surfaces worktrees in the UI: you can run three sessions on the same repo simultaneously, each on a different branch, and only converge at PR time. The CLI exposes the same primitive via --worktree-base-ref.
For agencies running Claude Code across multiple client repos, this removes a real category of bugs: lost changes when two sessions race the same working tree.
4. Auto memory keeps build commands and debugging insights
Auto memory was a snoozed idea in our backlog through May. At Tokyo it landed as a default Claude Code behavior. As Claude works on your project, it writes structured notes across three categories: project patterns (build commands, test conventions), debugging insights (solutions to tricky problems), and architecture notes (key files and module relationships). On the next session, those notes load automatically.
This is the Claude Code answer to the "I have to re-explain my project every session" complaint. It is also the foundation under Routines: a routine that runs nightly needs persistent memory of what worked last night.
5. Voice Dictation for hands-off prompting
Voice Dictation is a new audio input mode for Claude Code, surfaced both in the Desktop app and in some CLI configurations. The "Stop babysitting your agents" workshop at Tokyo demoed it as the fast path to start a long-running job: hold a hot key, describe the task in 20 seconds, hit Auto Mode, walk away.
This is not a flagship feature on its own, but combined with Remote Control it changes the workflow shape: you can start a Claude Code session from your phone in the morning, talk to it for 30 seconds, and come back to a finished PR at lunch.
6. Remote Control wakes sessions from your phone
Remote Control lets you monitor and operate long-running Claude Code tasks from a mobile browser. You can check what an agent is doing, send a follow-up instruction, approve a sensitive action, or kill a runaway session, all from a phone.
This is the feature that makes the "set it and forget it" framing real. A migration that takes 4 hours used to mean a developer stayed at the desk; now it means a developer kicks it off, opens the mobile UI when a Slack notification fires, and intervenes only if needed.
7. Routines: Claude Code on a cron
Routines are the biggest platform-level shift announced at Tokyo. A Routine is a saved Claude Code configuration (prompt, repos, environment, connectors) with one or more triggers attached:
- Scheduled: hourly, nightly, weekly, or a custom cron (one-hour minimum).
- API: HTTPS POST to a per-routine
/fireendpoint with a bearer token. - GitHub: react to
pull_request.opened,release.published, and similar repo events.
Routines run on Anthropic-managed cloud infrastructure, so they keep working when your laptop is closed. Anthropic ships Routines as a research preview behind the experimental-cc-routine-2026-04-01 beta header on Pro, Max, Team, and Enterprise plans with Claude Code on the web enabled.
Routine use cases Anthropic showcased at Tokyo:
- Backlog maintenance: every weeknight, triage new issues, label them, assign owners, post a Slack digest.
- Alert triage: a monitoring tool POSTs to the routine when an error threshold trips. Claude pulls the stack trace, correlates with recent commits, opens a draft PR with a proposed fix.
- Bespoke code review: on
pull_request.opened, apply your team's review checklist, leave inline comments, summarize for human reviewers. - Deploy verification: CD pipeline POSTs to the routine after each deploy. Claude runs smoke checks, scans error logs, posts go or no-go.
- Library port: on a merged PR in one SDK, port the change to a parallel SDK in another language.
This is the closest thing yet to "Claude Code as a cloud worker." It is also the most direct overlap with where Totalum's MCP path already lives: an agent that wakes up on a trigger, calls Totalum via MCP, builds or maintains a real app, and notifies you only when it is done.
8. Dynamic Workflows orchestrate dozens to hundreds of sub-agents
Dynamic Workflows is the orchestration layer surfaced inside a session. Inside one Claude Code session you can spawn dozens to hundreds of sub-agents working in parallel, each on a slice of the work. The Desktop app exposes a tasks pane that visualizes what each sub-agent is doing.
Tokyo demoed a fictional F1 team example: four agents running simultaneously analyzed aerodynamics, tire temperature, power units, and driver safety, each receiving evaluation from a grader agent and progressing autonomously. That is the same primitive scaled to enterprise workloads: a triage routine fans out to 50 sub-agents to look at 50 alerts in parallel.
9. Agent Teams for CLI-level parallelism
Agent Teams is the CLI-side counterpart to Dynamic Workflows. It is a set of parallel Claude Code sessions that can message each other inside the terminal. It is available in the CLI, not in Desktop (where the same problem is solved by Dynamic Workflows). For terminal-first teams, Agent Teams is the "no UI required" path to multi-agent work.
10. Managed Agents Vault for secret-safe automation
The Claude Managed Agents Vault, announced for general availability alongside Fable 5 on June 9 and recapped at Tokyo, stores environment variables and API credentials so a routine or a managed agent can call services without ever exposing the raw secret in a prompt or a log. Vault entries are scoped per environment and per routine.
This is the unglamorous but necessary piece. If you run Claude Code as a routine that posts to Slack, files Linear tickets, and pushes to GitHub, you need a place for those tokens that is not "paste them into the prompt." Vault is that place.
11. Security Scanning runs at night
Security Scanning is an automated, nightly job that crawls a repo for vulnerabilities and opens proposed-fix PRs. It is positioned as a default capability of a connected Claude Code account, not a separate product. Tokyo framed it as the obvious first Routine for any team: nightly schedule, repo input, Claude GitHub App connector, security-focused prompt, auto-merge gated on green CI.
Tokyo Day 1 Claude Code updates at a glance
| Feature | What it does | Status | Where it lives |
|---|---|---|---|
| Claude Code Desktop | Code tab inside Claude Desktop with per-session worktrees and panes | Generally available | macOS, Windows |
| Auto Mode | Classifier auto-approves safe tool calls; stops on risky ones | Generally available on Max, Team, Enterprise, API | CLI and Desktop |
| Worktrees | Each session gets its own isolated Git worktree | Generally available | CLI and Desktop |
| Auto memory | Persistent notes on project, debugging, architecture | Generally available | CLI and Desktop |
| Voice Dictation | Audio input for prompts | Generally available | Desktop and CLI |
| Remote Control | Monitor and operate sessions from a phone | Generally available | Mobile browser |
| Routines | Scheduled, API, or GitHub-triggered cloud sessions | Research preview | Web, CLI, Desktop |
| Dynamic Workflows | Spawn parallel sub-agents inside one session | Generally available | Desktop and CLI |
| Agent Teams | Parallel sessions that message each other | Generally available | CLI |
| Managed Agents Vault | Secure storage for routine secrets | Generally available | Claude Platform |
| Security Scanning | Nightly repo vulnerability scan with PR fixes | Generally available | Claude GitHub App |
What this means if you ship production apps
A few patterns are now reachable that were not a month ago.
Unattended fleet at agency scale. An agency that runs Claude Code across 15 client repos can stop pinning a developer to each long-running task. Convert the work into Routines, give each Routine a Claude GitHub App connection, and let the cloud run nightly. The agency's developers reserve their hours for the work the classifier escalates.
Phone-shaped builder loop. Voice Dictation plus Remote Control means a founder can start, monitor, and finish Claude Code work from a phone. Combined with the Claude Code skills marketplace, the loop becomes: open Claude on your phone, dictate the task, attach the relevant Skill, walk away.
Sub-agent orchestration without bespoke code. Before Dynamic Workflows, running 30 Claude Code instances on 30 slices of a repo required writing your own orchestrator. Now the orchestrator is a built-in primitive, and the Agent SDK sits on top of it for programmatic access.
Production app build via agent. This is the most interesting pattern for Totalum users. A Routine wakes up on a GitHub event, calls Totalum via MCP, asks Totalum to spin up a real Next.js project with auth, database, payments, and a custom domain, then writes a Slack note when the project is live. The agent is the orchestrator. Totalum is the AI app builder that materializes the project. The two compose cleanly.
How Totalum fits into the new Claude Code shape
Totalum is the AI app builder that creates real, production-grade Next.js applications with built-in auth, payments, database, file storage, AI integrations, deployment, and custom domains. It is a peer to Lovable, Bolt, Replit, and v0, not a hosting layer.
The Tokyo updates make the agent integration story bigger, not smaller. Two specific patterns:
- Claude Code (Desktop or Routine) drives Totalum via MCP. Connect the Totalum MCP server to your Claude Code account. The agent, whether running interactively in the Desktop app or unattended in a Routine, can now ask Totalum to create a new project, add a table, generate a page, deploy to a custom domain, or read back logs. The agent is the prompter. Totalum is the builder.
- SaaS founders embed Totalum behind their own AI builder feature. If you are building an AI app builder feature inside your own SaaS, the Routines pattern gives you a way to trigger long-running Claude Code work in the background while Totalum handles the actual project creation. Talk to us at Calendly if that is the pattern you want to ship.
If you are picking Claude Code over Cursor (or running both), our Cursor vs Claude Code 2026 verdict is still the cleanest comparison; the Tokyo updates push the verdict further in Claude Code's favor for unattended, long-horizon work. For the broader landscape, see best AI coding agents 2026.
What Day 1 did NOT ship
A few things builders were watching for did not appear at Tokyo Day 1:
- No new pricing model announcement. Programmatic Claude pricing is still scheduled to take effect on June 15, with the full credit pool and opt-in details covered in our Claude Agent SDK credits 2026 breakdown.
- No live Partner Hub launch. The Partner Hub framework was confirmed at the Claude Partner Network announcement on June 3, but the live directory was not shown at Tokyo Day 1.
- No Memory Files GA. The Memory Files preview from earlier in 2026 stayed in preview through Tokyo.
- No public Mythos branding launch. Mythos 5 stays gated behind Project Glasswing as announced on June 9. See our Claude Fable 5 and Mythos 5 deep dive for the model split.
Tokyo Day 2 Extended on June 11 is a community day. We do not expect product news from it.
FAQ
When was Code with Claude Tokyo 2026 held?
Code with Claude Tokyo Day 1 was held on June 10, 2026, in Tokyo. Code with Claude Extended (Day 2) ran on June 11, 2026, targeting independent developers and early-stage founders. The event followed the San Francisco edition (May 6, 2026) and the London edition (May 19, 2026).
What is the biggest announcement from Code with Claude Tokyo Day 1?
The biggest shift is the packaging: Anthropic positioned Claude Code as a platform rather than a CLI. The single most consequential individual launch was Routines, which lets a Claude Code session run on a cron, react to a GitHub event, or trigger from an HTTP POST. Routines are the closest thing yet to "Claude Code as a cloud worker."
Is Claude Code Desktop generally available?
Yes. Claude Code Desktop is in general availability for macOS (Intel and Apple Silicon) and Windows (x64 and ARM64). Linux users continue to use the CLI; there is no native desktop app on Linux as of June 10, 2026.
Are Routines free?
Routines draw down the same subscription usage as interactive Claude Code sessions, plus a daily cap on routine runs per account. One-off routines (single future run) do not count against the daily cap. Routines are available on Pro, Max, Team, and Enterprise plans with Claude Code on the web enabled. Team and Enterprise admins can disable Routines for all members.
Can I run Claude Code on my phone now?
Sort of. Remote Control gives you a mobile browser UI to monitor, operate, and approve sessions started on your laptop or in the cloud. Voice Dictation gives you audio input for prompts. Both together get you close to "start a Claude Code job from your phone, walk away, and finish on a laptop later." There is no native phone app for Claude Code yet.
How does the Tokyo platform shift affect Cursor vs Claude Code?
Cursor still leads on in-IDE design work and on Composer 2.5 latency for in-flight edits. Claude Code now leads on unattended long-horizon work, parallel sub-agent orchestration, and triggered cloud automation. For most agencies running both, the split is: Cursor in the IDE for hands-on work, Claude Code as a Routine for nightly maintenance and incident triage.
How does Claude Code Routines compare to Totalum MCP?
They compose. A Routine is a trigger and an orchestration shell for Claude Code. Totalum MCP is the build target that turns "the agent wants a real Next.js app with auth, payments, database, and a custom domain" into an actual deployed project. A Routine that needs to create or maintain a real app should call Totalum via MCP, rather than try to write the app from scratch each time.
Ready to build with Totalum?
If you run an agency or a SaaS team and want to combine Claude Code Routines with Totalum's AI app builder via MCP or API, book a 30-minute call at calendly.com/totalum/30min. We will walk through the exact integration shape.
If you are a Claude Code power user who wants to drive Totalum from your existing agent loop, start free at totalum.app and connect via MCP from your Claude Code Desktop or Routine.
Sources: Anthropic news, Code with Claude Tokyo, Claude Code Routines docs, Claude Code Desktop docs, Claude Platform release notes.