Computer use, Skills API, and Files API go generally available, and Python SDK v1.0 lands with breaking changes

TLDR;

Computer use, the Skills API, and the Files API all hit general availability on the Claude Platform this week: the beta headers are gone and production use is officially supported. The Anthropic Python SDK v1.0 also dropped with breaking changes that require migration; run /claude-api upgrade in Claude Code v2.1.239+ to automate the Python-side work. MCP published its new roadmap and Claude Code shipped eight releases including Remote Control going generally available and the new /design artboard skill.

Headline

Computer Use, Skills API, and Files API are Generally Available

As of August 20, the computer_toolset_20260801 toolset requires no beta header, adds batch actions and zoom-by-default, and supports per-member configs. The Skills API (/v1/skills) and Files API are also fully generally available: upload skills once and reference file IDs across requests instead of re-sending payloads. Builders on existing computer-use integrations should review the Migrate from computer_20251124 guide, as the request shape and tool-handling have changed.

1. API & Model Updates

Anthropic Python SDK v1.0: Breaking Changes and Migration Required

The Python SDK moved from httpx to httpx2, dropped Python <3.10, removed the legacy Text Completions API, eliminated temperature/top_p/top_k from Messages method signatures, and now raises an error (instead of silently defaulting) when AnthropicBedrock has no AWS region configured. On the async client, .with_raw_response results now require await response.parse(). Run claude /claude-api upgrade (requires Claude Code v2.1.239+) to automate migration of Python projects from anthropic 0.x to 1.x.

Admin API User Management is Generally Available (August 18, 2026)

The Admin API user-management endpoints for Claude Enterprise (members, invites, groups, and custom roles) are out of beta as of August 18. The anthropic-beta: ce-user-management-2026-07-13 header is no longer required on group and custom-role requests; requests that still send it continue to work.

Web Search and Web Fetch Domain Restrictions Now Available (August 18, 2026)

You can now restrict which sites a Claude Managed Agents agent's web_search and web_fetch tools can reach by setting allowed_domains or blocked_domains on the tool entry in the agent_toolset_20260401 configs array. web_fetch also gains max_content_tokens and web_search gains user_location.

Legacy Claude API Workbench Retired (August 17, 2026)

The legacy Claude API Workbench was shut down on August 17 with no recovery path for missed exports: saved prompts, variables, and evals are gone. Three experimental prompt-engineering endpoints now return errors; each can be replicated with a standard Messages API meta-prompt.

2. Skill of the Week

ELI5, the explainer skill Anthropic uses internally

This week's editor pick. Thariq (@trq212) of Anthropic shared the skill Anthropic staff have been using a lot internally (Aug 21): type /eli5 and Claude explains it "like I'm someone who knows nothing about this topic, using a HTML artifact with big pictures and few words." The output is a clean HTML artifact: big visuals, simple diagrams, almost no walls of text. The post passed 13,000 likes within a day, and the skill itself is a one-line prompt you can recreate as a custom skill in seconds. A good reminder that the highest-leverage skills are often prompt one-liners, not code.

3. Tool of the Week

fx by Vercel

Vercel released fx (Aug 18), a tiny, open, native coding agent written in Zig. The entire agent is a 6.13 MiB binary that cold-starts in about 10 microseconds, keeps a single-digit-megabyte memory baseline, and ships WebAssembly builds for embedding in browsers and sandboxes. It is model- and provider-agnostic (local models, gateways, direct API access, or subscriptions), Apache-2.0 licensed, and extends via skills, plugins, and MCP servers. The form factor deliberately reads like a Unix shell rather than an "IDE in the terminal" TUI, and it is optimized for research and embedding inside larger agent systems. Still experimental (v0.0.5, frequent changes expected), but the vercel-labs/fx repo passed 2,200 stars in its first week. Try it: curl -fsSL https://fx.sh/setup.sh | bash.

4. Claude Code & CLI

`/design` Skill, Remote Control Generally Available, and Concise Output Style

Week 34's official digest (covering v2.1.234–v2.1.239) names three headlining features: the /design artboard skill (research preview, requires v2.1.233+: give it a brief and Claude publishes a canvas of editable artboards for your UI, iterate in natural language, then have Claude implement the one you pick in code); Remote Control out of research preview, meaning any machine running claude remote-control now appears as a device card in the Code tab on your phone, letting you start a session remotely; and the built-in Concise output style, which leads with results and skips preamble while doing work as thoroughly as Default (selectable under Output style in /config, ships in v2.1.237).

Sessions Auto-Resume at Usage-Limit Reset

Claude Code now continues your session automatically when a claude.ai usage limit resets, removing the need to manually restart. Toggle the behavior off in /config ("Continue automatically at usage limit").

`ANTHROPIC_DEFAULT_MODEL` Env Var Added

The new ANTHROPIC_DEFAULT_MODEL environment variable sets the model new sessions start on, while a /model pick still overrides it and persists across restarts, unlike ANTHROPIC_MODEL, which locks the model. Useful for CI/CD pipelines and team configs where you want a default without preventing developers from switching.

Dynamic Workflows, `/effort ultracode`, and Subagent Fan-Out

The /workflows command lets Claude write a reusable orchestration script that fans out work to parallel subagents, suited for codebase audits, large migrations, and cross-checked research. Pair with /effort ultracode for maximum reasoning and orchestration depth. Scripts are saved and re-runnable without re-prompting.

How Anthropic Contains Claude Across Products

Published August 20, this Anthropic engineering post covers the mechanisms they use to contain Claude's behavior across their product surface: relevant context for builders building on top of the platform who need to understand where Anthropic's guardrails sit vs. where operator-level controls apply.

5. MCP Ecosystem

MCP Publishes New Roadmap (August 22, 2026)

The Model Context Protocol roadmap was updated August 22 by Lead Maintainers David Soria Parra and Den Delimarsky, covering the next spec release and a 6–12 month horizon. Five priority areas: (1) Agentic messaging primitives: server-initiated events (webhooks/channels) and a composition review across Tasks, Transports, and Triggers & Events working groups, with SEP-2663 (Tasks extension) moving toward core-spec inclusion; (2) HTTP-native transport unification: HTTP/2 over stdio as a single binding, plus ETags for caching tool-call results; (3) Agent identity and enterprise-ready security: authorization for agent-to-agent flows where no human is present at consent time; (4) improved primitives; (5) improved SDK developer experience. The full roadmap includes Working Group assignments and SEP references for each area.

6. Competitor Dev Tools

Cursor Cloud Agents: Subscriptions, `/goal`, Subagents on Isolated VMs (August 19, 2026)

Cursor's August 19 release adds persistent goal-tracking (/goal), agents that subscribe to PR and Slack events and wake when something happens, and subagents running on isolated virtual machines with clean context. These features land in the same week as Claude Code's own subagent fan-out and /effort ultracode: builders evaluating both tools now have a closer feature-parity comparison to make.

7. Research & Safety

How Claude is Accelerating Protein Design and Analytical Chemistry

Published August 18, this Anthropic research post covers applied science use-cases for Claude in protein design and analytical chemistry workflows: useful context for builders targeting life-sciences verticals who want to understand what capability profiles Anthropic is actively developing and documenting.

Release Notes

Claude Code:

v2.1.241 (Aug 23): Bug fixes and reliability improvements
v2.1.240 (Aug 22): Bug fixes and reliability improvements
v2.1.239 (Aug 21): Adds /claude-api upgrade for 0.x→1.x Python SDK migration; cost estimates now include 1.1× US-only inference premium; fullscreen renderer offer on Bedrock/Vertex/Foundry; fixes Bedrock streaming proxy bug that silently doubled billed API calls (security-relevant); fixes SSO+HTTPS proxy startup hang; fixes Edit/Write 5-second pause in JetBrains
v2.1.238 (Aug 20): Adds keybindingFlavor: "readline" for Bash-style Ctrl+W; plugin headersHelper for token-minting on marketplace fetches; --defer-shutdown-max-min for self-hosted runners; fixes unbounded memory growth in long sessions; fixes output styles drifting back to default voice mid-session
v2.1.237 (Aug 20): Fixes prompt caching for LLM gateway / custom base URL sessions; adds built-in Concise output style
v2.1.236 (Aug 19): Adds ANTHROPIC_DEFAULT_MODEL env var; cross-session SendMessage gains one-shot idle notifications; fixes sandbox wildcard read-deny bypass via file rename (security-relevant); fixes clipboard, fullscreen renderer, and /model picker regressions
v2.1.235 (Aug 18): Adds optional spellcheck setting (requires aspell/hunspell/ispell); fixes Shift+Tab in permission comment field prematurely approving edits; fixes whole-prompt-cache invalidation on language server reconnect; improves memory/CPU for cloud sessions
v2.1.234 (Aug 17): Remote file reads, session restore, CLAUDE.md includes, and workflow scripts now reject Windows NT-namespace paths, closing the NTLM credential-leak vector (security-relevant); sessions auto-resume at usage-limit reset; adds GitLab MR badge in footer

Agent SDKs:

TypeScript v0.3.241 (Aug 23): Parity with Claude Code v2.1.241
TypeScript v0.3.240 (Aug 22): Parity with Claude Code v2.1.240
TypeScript v0.3.239 (Aug 21): total_cost_usd now includes 1.1× data-residency multiplier; fixes SYSTEM_PROMPT_DYNAMIC_BOUNDARY being sent as literal text on Bedrock/Vertex/Foundry/gateway providers; reconnecting hosts now receive background_tasks_changed snapshot
TypeScript v0.3.238 (Aug 20): Adds is_backgrounded and spawn_depth to task_started events; adds suppressOriginalPrompt to UserPromptExpansion; fixes SDK hooks silently not applying after host re-sends initialize
TypeScript v0.3.237 (Aug 20): Parity with Claude Code v2.1.237
TypeScript v0.3.236 (Aug 19): PostToolUse hooks can now return hookSpecificOutput.classifierContext to annotate tool-call results for the auto-mode permission classifier
TypeScript v0.3.235 (Aug 18): Parity with Claude Code v2.1.235
TypeScript v0.3.234 (Aug 17): Removes unused bypass_permissions_disabled from ExitReason (compile error on upgrade if you have an explicit case branch; runtime unaffected); SDKSystemMessage gains effort field; vcs_state_changed reports shell's final directory
Python v0.2.143 (Aug 20): Bundles Claude CLI v2.1.238
Python v0.2.142 (Aug 20): Bundles Claude CLI v2.1.237
Python v0.2.141 (Aug 19): Bundles Claude CLI v2.1.236
Python v0.2.140 (Aug 18): MCP 2.x support for in-process SDK MCP servers (mcp>=1.23.0,<3.0.0); adds forward_subagent_text option; adds ResultError exception with structured error payload; can_use_tool callback now works with string prompts

Action Items

Immediate:

Migrate off Anthropic Python SDK 0.x: v1.0 is out with breaking changes (httpx→httpx2, Python 3.10 minimum, Text Completions removed, Bedrock region now required). Run claude /claude-api upgrade in Claude Code v2.1.239+ to automate Python project migration. If you're not ready, pin anthropic<1.0.0 explicitly now.
Remove legacy Workbench API calls: three experimental prompt-engineering endpoints now return errors as of August 17. Rebuild against the standard Messages API; there is no recovery path for missed prompt exports.
Audit Claude Code deployments for the NTLM credential-leak fix: v2.1.234 closes a pre-approval file-access vector via Windows NT-namespace paths. Update to v2.1.234+ immediately in any Windows or mixed-OS environment.
Check TypeScript Agent SDK v0.3.234 upgrade: if you have an explicit case branch on bypass_permissions_disabled in your ExitReason handling, you will get a compile error on upgrade. Remove the branch before upgrading; runtime behavior is unaffected.

By August 31, 2026:

Audit for computer_20251124 beta header usage: the generally available toolset is computer_toolset_20260801; the request shape and tool handling have changed. Review the Migrate from computer_20251124 guide linked in the general-availability announcement.
Pin MCP Python SDK version if not ready for 2.x: the 1.x line moves to security-fixes-only maintenance mode. Pin mcp>=1.28,<2 to avoid surprise breakage, then plan your 2.x migration.

Want to catch up? Read every past issue of This week in Claude.

Did you like this issue?

One tap is the fastest way to shape next week.

👍 Liked it · 👎 Didn't like it

And I read every reply. Hit reply and tell me what you want more of, or less of, and it goes straight into the next issue.

Mert Duzgun