Claude Code August 20: v2.1.237 Adds a Built-in Concise Output Style, the Files and Skills APIs Go GA, and the 50 Percent Weekly Limit Boost Runs to August 31

Anthropic API: Files and Skills APIs reach GA, computer use and browser use toolsets added

• August 19: the Files and Skills APIs moved to general availability in TypeScript SDK v0.119.0, which also adds computer use and browser use toolsets.

Editor’s read: GA is the signal to build on Skills rather than around them. If you have been holding a Skills integration behind a preview flag, this is the release that unblocks it.

Anthropic TypeScript SDK v0.120.0 adds managed agents web search config and self-hosted sandbox memory

• August 19: v0.120.0 exposes web search configuration for managed agents and memory settings for self-hosted sandboxes.

Anthropic TypeScript SDK v0.117.0: output_behavior on dream creation and a streamed message_delta accumulation fix

• August 13: v0.117.0 adds output_behavior on dream creation, fixes streamed message_delta accumulation and honors the per-request timeout.

Claude Code v2.1.237 ships a built-in Concise output style and fixes prompt caching behind an LLM gateway

• August 20: the new Concise output style leads with results and skips preamble and narration, selectable under Output style in /config.
• The same release fixes prompt caching for sessions running through an LLM gateway or custom base URL.

Editor’s read: If you have been maintaining a custom terseness prompt, delete it and use the built-in style instead.

Claude Code v2.1.236 adds ANTHROPIC_DEFAULT_MODEL, notify_when_idle and macOS sandbox read-deny hardening

• August 19: ANTHROPIC_DEFAULT_MODEL sets the model new sessions start on while a /model pick still overrides it and persists.
• Cross-session SendMessage gains an opt-in notify_when_idle.
• On macOS, wildcard read-deny rules such as **/.env now take precedence inside allowed read regions, cover matched directories, and cannot be bypassed by renaming the denied file.

Editor’s read: The sandbox change is the important one. If you relied on a read-deny rule to protect .env files, it now actually holds.

Claude Code v2.1.235 adds prompt-input spellcheck and fixes Shift+Tab granting session-wide edit permission

• August 18: an optional spellcheck setting underlines misspelled words in the prompt input using aspell, hunspell or ispell.
• A permission-prompt bug where Shift+Tab inside the comment field approved the edit and granted session-wide edit permission is fixed, and memory and CPU use during background cloud sessions is lower.

Claude Code v2.1.234 auto-continues when the usage limit resets and adds a GitLab MR badge

• August 17: sessions now auto-continue when the usage limit resets, a GitLab merge-request badge appears, and NT-namespace path handling was hardened.

Claude Code v2.1.233: GitLab merge-request worktrees and todo tools off by default on Opus 4.8 and Sonnet 5

• August 14: adds GitLab merge-request worktrees and Linux Bash memory limits, and turns todo tools off by default on Opus 4.8 and Sonnet 5.

Claude Code v2.1.232 turns subagent forking on by default and lets you @-mention another Claude session

• August 13: subagent forking is on by default, so a fork inherits the whole conversation and the prompt cache instead of starting cold.
• You can also @-mention another of your sessions by name.

Editor’s read: Forked subagents change the economics of delegation. A fork that inherits the prompt cache is markedly cheaper than a cold subagent.

Claude Code v2.1.231 fixes MCP OAuth sign-in failing with 'redirect URI mismatch'

• August 13: fixes MCP OAuth sign-in failing with the literal error 'redirect URI mismatch' on pre-registered clients such as Slack.

Claude Code v2.1.229 adds remote-control --continue and plugin marketplace command sources

• August 12: adds remote-control --continue and plugin marketplace command sources, and fixes Windows UNC path and narrow-terminal crashes.

Claude Code v2.1.228 hardens claude.ai-synced skills and fixes a Remote Control resume leak

• August 11: hardens claude.ai-synced skills, fixes a Remote Control resume leak and changes Write tool overwrite behavior.

Claude Code v2.1.227 fixes the Max plan usage-credits prompt and claude-code-action Bash on GitHub runners

• August 10: fixes the Max plan usage-credits prompt, claude-code-action Bash failures on GitHub runners, and slash-command handling.

Claude Agent SDK for TypeScript v0.3.236: PostToolUse hooks can return classifierContext

• August 19: PostToolUse hooks can now return hookSpecificOutput.classifierContext, a short host-asserted note about a tool call's result that the auto mode permission classifier reads alongside that result.

Editor’s read: This is the cleanest hook yet for teaching auto mode why a tool call was safe. Worth wiring if you run unattended agents.

Claude Agent SDK for Python v0.2.140 adds MCP 2.x in-process servers and structured ResultError

• August 18: adds MCP 2.x in-process servers, forward_subagent_text and a structured ResultError.

Claude Agent SDK for TypeScript v0.3.234 returns real ApiKeySource values and adds effort on init

• August 17: real ApiKeySource values, an effort field on init and ExitReason cleanup.

Claude Agent SDK for TypeScript v0.3.232 adds a structured context_usage payload

• August 13: adds the structured SDKContextUsage context_usage payload, _meta on subagent tool_use_result and branch on vcs_state_changed.

Claude Agent SDK for Python v0.2.137 adds ConversationResetMessage and resume_session_at forking

• August 12: adds ConversationResetMessage, message origin on UserMessage and ResultMessage, and resume_session_at forking.

Claude Agent SDK for TypeScript v0.3.229 surfaces terminal_slash_commands and ends oversized turns as api_error

• August 12: adds terminal_slash_commands to the init message; turns that hit the 32 MB request limit now end as api_error.

Claude Agent SDK for TypeScript v0.3.228 carries output_tokens_details through Agent tool results

• August 11: output_tokens_details is now carried through Agent tool results.

MCP reference servers 2026.8.18 updates the everything, time, fetch and git packages

• August 18: the reference server release refreshes server-everything, mcp-server-time, mcp-server-fetch and mcp-server-git.

Serverless Framework deploys MCP servers on AWS Lambda in four lines of YAML

• August 11: Daniel Avila showed the Serverless Framework deploying an MCP server to AWS Lambda with four lines of YAML.

Theo on the new MCP version: genuinely good, with a significant catch

• August 11: Theo called the new MCP version genuinely really good but flagged a substantial caveat that comes with it.

Claude Code usage limits: the 50 percent weekly increase is extended only through August 31

• August 18: Anthropic extended the 50 percent increase to weekly Claude Code limits through August 31, saying it hopes to make the change permanent but that strong demand means capacity may be tight.

Editor’s read: This is the fourth short extension rather than a permanent change. Plan your August capacity on the assumption it can lapse.

Claude Code named sessions can now message each other from the terminal

• August 13: launch a session with claude --name backend, then have another session message it by name.

Claude in Chrome sessions now carry over to desktop, web and mobile with skills and connectors

• August 12: Claude in Chrome sessions continue across desktop, web and mobile, carrying skills and connectors with them.

Theo spent a week on Matt Pocock and PStack skills: do not copy them wholesale

• August 19: a widely shared writeup of Theo's 38-minute review.
• The most reusable takeaway is that the SKILL.md description field is misnamed and should be thought of as a trigger, because the model only sees the name and description when deciding whether to pull a skill into context.

Editor’s read: If your skills fire at the wrong time, rewrite the description as a trigger condition rather than a summary of what the skill does.

Matt Pocock: /improve-codebase-architecture is the rescue hatch for a broken vibe-coded app

• August 10: Pocock recommends the improve-codebase-architecture skill when a vibe-coded application starts falling apart.

Matt Pocock on shared codebase language as a lever on agent output quality

• August 12: learning and using your codebase's own shared vocabulary measurably improves what coding agents return.

Matt Pocock's /to-tickets tip: size tasks by estimated token spend before splitting

• August 11: estimate token spend per task and use that to decide where to split work into tickets.

Claude Code creator Boris Cherny: LLM bugs have moved from off-by-ones to system design

• August 11: Cherny observes that the bugs models produce have shifted away from off-by-one errors toward system design and usability problems.

Z.ai GLM-5.3 targets Claude Code and OpenAI Codex users

• August 14: Z.ai says GLM-5.3 codes 50 percent better than GLM-5.2 and tops Terminal Bench 3.0 among open models, with weights to follow.

Devin Cloud Agents run macOS, Linux, Windows and Android sandboxes in parallel

• August 12: Nader Dabit highlighted Devin Cloud Agents running parallel sandboxes across four operating systems.

Claude Code pricing analysis: Max 20x at $200 per month against roughly $8,000 in API-equivalent usage

• August 11: an enterprise pricing breakdown argues the same tokens on the same model can cost up to 40 times less on a Max 20x subscription at $200 per month than the API-equivalent spend of about $8,000.

Editor’s read: Treat the 40x figure as a ceiling from heavy usage rather than a typical result, but the direction is right: subscription seats beat API billing for sustained interactive work.

Anthropic will watermark all Claude text output, covering Claude Code, the API and Cowork

• August 11: Anthropic announced that models launched on or after August 2, 2026 embed statistical patterns into Claude's text output and add C2PA metadata to files such as images, framed as meeting EU AI Act transparency commitments.
• Follow-up coverage on August 12 confirmed the rollout spans Claude Code, the API and Cowork worldwide.

Editor’s read: If you ship Claude-generated text in a product, assume it is detectable. Factor that into your disclosure policy now rather than after a customer asks.

Anthropic Vertex and Bedrock TypeScript SDKs patched so ambient credentials never reach provider endpoints

• August 13: vertex-sdk 0.19.2 and bedrock-sdk 0.32.2 ensure ambient first-party credentials are never sent to provider endpoints.

Editor’s read: This is a credential-leak class fix. If you run either SDK in a cloud environment with ambient credentials, upgrade rather than schedule it.