Blog

The Admin API landed in the ant CLI and seven language SDKs on 26 August 2026 under client.beta.organization. Here's what it covers, what still needs curl, and how to authenticate it.

Anthropic's own guidance for defining custom tools on the Claude API and Claude Managed Agents comes down to four rules, and the first one is the same lever that makes or breaks an agent skill.

Claude Fable 5.1 shipped 1 September 2026 as the default Fable model, with a quarter-price cache read and three breaking changes for anyone calling Claude Fable 5 today.

Codex CLI's hooks system gained async execution, MCP tool access, an Interrupt event, tool-result rewriting and per-tool output limits across five releases. Here's every change in order.

Claude's browser use tool drives a webpage through its accessibility tree and element references, not just screenshots. Here's how it works and when to reach for it over computer use.

The claude-api skill's cost-optimize subcommand profiles a project's Claude API spend and proposes savings, caching first, one measured change at a time, instead of generic advice.

Claude Managed Agents' web_search and web_fetch tools now accept allowed_domains and blocked_domains, each with its own list. Here's the exact config shape and validation rules.

modelPicker lets you replace or extend Claude Code's built-in /model picker with your own ordered, labelled list of models, ideal for teams standardising on specific Bedrock or Vertex deployments.

Anthropic confirms Claude Sonnet 5's launch pricing of $2/$10 per million tokens is now permanent. The scheduled rise to $3/$15 on 1 September 2026 will not happen.

Anthropic took Agent Skills and the Skills API out of beta on the Claude API on 19 August 2026. What GA actually changes, what the skills-2025-10-02 beta header still does, and the SDK migration that followed.

Claude Code v2.1.248 brought cross-session messaging to Amazon Bedrock, Google Cloud's Agent Platform, Microsoft Foundry, and sessions with telemetry disabled, providers that were explicitly excluded until now.

Claude Code v2.1.251 added PreModelSwitch and PostModelSwitch hook events for blocking, confirming, or reacting to a model switch. How they fire, what they receive, and how blocking actually works.

Claude Code v2.1.251 streams a foreground subagent's tool calls and results live to Remote Control clients, while background subagents still show status only. What changed and why it only covers one kind of subagent.

How to read your organization's Claude API rate limits programmatically, and how workload identity federation endpoints let a CI workload manage its own config without a long-lived key.

How the experimental.cacheTtl field in subagent frontmatter sets a per-agent prompt cache lifetime, where it sits in the precedence order, and when it actually changes anything.

How the managed-only modelPricing setting makes /usage, the status line and telemetry report spend at your contracted rates instead of list price, without changing your bill.

Claude Code v2.1.251 fixed a symlink escape in Read, Write and Edit, a plugin command path-traversal bug, a Workflow tool permission bypass and more. What each fix actually changed.

Custom instructions load on every message and stay short by necessity. Agent Skills load only when relevant and can be as long and detailed as a procedure needs. Here's when to use each.

Members, invites, workspaces, RBAC groups and custom roles: what the Claude Admin API actually covers for managing an organization programmatically, and where Enterprise-only endpoints start.

A beta that lets you add or withdraw tools partway through a Claude API conversation without touching the tools array, so the prompt cache stays intact.

Claude Code's new --restricted flag locks a session down for evaluation harnesses and shared machines: no shell tools, no WebFetch, no user or project settings, no bypass mode.

Claude Code v2.1.246 added an Auto mode tab to /permissions so you can view and edit the classifier's allow, soft_deny, hard_deny and environment rules without opening a settings file.

Claude Code v2.1.247 added a SendFeedback tool that drafts a feedback report the moment something goes wrong in a session, for you to review, edit or discard from /feedback.

Claude Code v2.1.243 added a /status line showing whether GitHub is connected for Claude Code on the web, plus a tip pointing you to /web-setup. Here is what each connection method actually does.

Claude Managed Agents can run tool execution in an Anthropic-managed cloud sandbox or on your own infrastructure. Here is what a self-hosted sandbox actually gives up, and what it buys you.

How to version-control Claude API resources as YAML with the ant CLI, chain commands with --transform, inspect errors, and call ant from inside Claude Code.

How the Claude Messages API's advisor tool pairs a cheap executor model with a stronger advisor for mid-generation guidance: setup, result variants, cost and model compatibility.

How Claude Code's promptCacheTtl and subagentPromptCacheTtl settings control the API's 5-minute and 1-hour cache lifetimes, who gets which default, and when to change it.

How Claude Managed Agents session and deployment budgets cap spend at public list rates: max_list_cost, the budget_reached stop reason, and multiagent sharing.

ant is Anthropic's official command-line tool for the Claude API, covering install, authentication, the resource:action command pattern, and how it differs from curl and the SDKs.

How a Claude Managed Agents coordinator delegates to a roster of other agents: session threads, the advisor pattern, MCP server scoping, and the hard limits on delegation.

How Scheduled Deployments run a Claude Managed Agent on a recurring cron schedule: creating one, budgets, run history, pausing, and the DST edge cases to watch for.

The anthropic Python SDK hit 1.0.0 on 20 August 2026, moving to httpx2 and dropping several long-deprecated APIs. Here is every breaking change and how to work through them.

Fullscreen rendering removes flicker, keeps memory flat and adds mouse support in Claude Code. Here is how to turn it on, what breaks, and why Bedrock and Vertex only just got it.

Claude Managed Agents is Anthropic's hosted agent harness: versioned agent configs, sandboxed sessions, and skills loaded from a GitHub repo. Here is how it differs from the Messages API.

Codex CLI 0.148.0 added codex exec fork for headless session branching, /export to Markdown, session archiving, thread cost visibility, Bedrock as a provider and async hooks.

The claude-api skill ships bundled with Claude Code and installs anywhere Agent Skills work, giving Claude current reference material for the Messages API and Claude Managed Agents.

Claude Code v2.1.239 made /cost, the status line, and --max-budget-usd account for the 1.1x US-only-inference premium on data-residency workspaces. Here's what that premium is.

Claude Code and Cursor both read the same SKILL.md standard, but one is a CLI and the other an IDE. Here's how install paths, activation, and team workflows differ between them.

Codex CLI 0.149.0 added a codex agents dashboard for managing tasks, plus /cd, /pwd and /cwd, codex queue, expanded Vim editing and a more capable codex doctor. Here's what actually changed.

ANTHROPIC_DEFAULT_MODEL sets which model a new Claude Code session opens on, without touching settings.json, but six other sources can override it. Here's the exact precedence.

Claude Code v2.1.234 added automatic continuation when a session or usage limit resets, turned on by default and controlled from /config, plus a separate Desktop app checkbox with its own mechanics.

Claude Code reads keyboard shortcuts from ~/.claude/keybindings.json, with contexts, chords, and unbinding, plus a separate keybindingFlavor setting for readline-style word editing.

notify_when_idle lets one Claude Code session ask another on the same machine to send one notice the moment it finishes, replacing manual polling with a one-shot, opt-in subscription.

Eight MCP servers worth connecting to a coding agent in 2026, chosen because each is maintained by the company whose name is on it, not a third-party wrapper around an undocumented API.

Claude Enterprise organisations can now turn on automated scanning that checks custom skills for malicious behaviour before members can use them. Here is exactly what it covers and what it doesn't.

MCP's Tasks extension lets a server hand back a task ID instead of blocking on a slow tool call, so a client can poll for the result later. Here is the actual mechanics: states, methods and the polling flow.

Output styles change how Claude Code communicates by rewriting its system prompt. Four built-in options now exist, including the new Concise style added 20 August 2026, plus how to write your own.

Claude Code's skill listing has a budget scaled to about 1% of context. Once it overflows, your least-invoked skills lose their descriptions first, a common reason a skill quietly stops triggering.

skillOverrides hides or restricts a skill from Claude Code settings instead of editing its SKILL.md, with four states from fully listed to completely hidden, written for you by the /skills menu.

The Model Context Protocol's 2026-07-28 revision drops the stateful handshake for a stateless core, adds mid-call user input, and deprecates Roots, Sampling, Logging and the old HTTP+SSE transport.

Channels push messages, CI results and webhooks straight into a running Claude Code session so it can react while you're away, instead of Claude polling for updates on a timer.

Monitor lets Claude Code watch a log, a CI job, a directory or a WebSocket feed in the background and react the moment something changes, instead of you re-prompting on a timer.

Subagent forking lets a subagent inherit the full parent conversation and prompt cache instead of starting fresh, and it now runs in the background by default. Here's exactly how it works.

Seeing a skill trigger only proves Claude found it, not that it did the right thing. Here's how to test both discovery and output quality before you publish a SKILL.md.

Agent teams let several Claude Code sessions coordinate through a shared task list and direct messages, with one acting as lead. Here's how it actually works, and where it's still experimental.

/goal sets a completion condition and keeps Claude working across turns until a separate evaluator model confirms it's met. Here's exactly how the evaluation loop works, and when to reach for it instead of /loop.

Claude Code now has three ways to run on a schedule: cloud Routines, Desktop scheduled tasks and session-scoped /loop. Here's what each is actually for.

In two releases on 13 and 14 August 2026, Claude Code added GitLab merge request support to --worktree and agent view, secret redaction for GitLab tokens, and GitLab plugin marketplaces.

GitHub Copilot has four individual tiers as of August 2026, Free, Pro, Pro+ and Max. Here's exactly what each includes, and which one you need for agent skills and third-party coding agents.

The quickstart gets one runner talking to Anthropic. This guide covers what production actually needs: egress control, git credentials, Kubernetes and Compose recipes, and drain timing.

A practical path through Claude Code for someone who has never opened a terminal AI agent before: install, first session, the commands worth memorising, and what to learn next.

Auto mode's classifier only trusts your working repo by default. Here's how to use autoMode.environment to add your org's repos, buckets and services, and stop routine work from getting blocked.

OpenAI's Codex CLI 0.147.0 extended its /import command to pull in Cursor-managed skills, not just Claude Code's. Here's what actually moves, and what to check before you trust it.

Claude Code sessions can now message each other with ListAgents and SendMessage. How discovery, delivery, and inbound controls work, and how it differs from subagents.

Claude Code's self-hosted environments run cloud sessions on your own infrastructure. Here's how to set up the runner, route a session to it, and what changes for skills.

From 14 August 2026, auto mode replaces manual permission prompts by default for new Claude Code sessions on Pro, Max and Team. Here's what the classifier actually allows and blocks.

How to install Agent Skills in the Zed code editor: the skills CLI, manual setup in .agents/skills, global versus project scope, worktree trust, and troubleshooting.

A new open standard from OpenAI, Amazon, Microsoft, Cursor and Vercel packages Agent Skills and MCP servers into one distributable plugin. Here's what it actually changes.

How to install Agent Skills in Google Gemini CLI: the native gemini skills command, manual setup in ~/.gemini/skills, the consent prompt, and troubleshooting.

Nine AI CRMs compared on what matters in 2026: what the AI actually costs once agents are metered, and which ones your coding agent can genuinely drive over MCP.

Nine AI image generators tested on what still separates them in 2026, in-image text, character consistency, commercial licensing and real API cost per image.

The standard shipped in late 2025. 2026 is when it spread, Cowork, Record a Skill, a cross-agent installer, and tens of thousands of public skills. Here's where things actually stand, opinions included.

A category-by-category guide to the agent skills worth exploring for operations, finance and content teams, what to look for, and where to find them.

A curated selection of the strongest skills for authoring, structuring and testing your own agent skills, plus how Record a Skill fits alongside hand-writing SKILL.md.

A curated selection of the strongest productivity skills for AI agents, task boards, invoice chasing, CRM upkeep and browser automation, picked for non-technical teams too.

A curated selection of the strongest marketing skills for AI agents (brand voice, planning, content, email and competitive research) picked for non-technical teams too.

A curated selection of the strongest research skills for AI agents, evidence mapping, literature review, trend scanning and document verification, with honest caveats.

A genuine comparison of Claude Cowork's Record a Skill against hand-written SKILL.md files, where each wins, concrete examples, and why the right answer is usually both.

Opus 5 shipped 24 July 2026 with real gains in agentic coding and multi-agent coordination. Here's what that actually means for how you write and structure skills.

How Agent Skills work in Claude Cowork, adding skill folders, the Team skill directory, Record a Skill, and how this GUI approach differs from Claude Code's filesystem model.

A full walkthrough of Record a Skill in Claude Cowork, where to find it, who has access, how the demonstration becomes a skill, sharing on Team, and privacy considerations.

Ten curated agent skills for copywriting in 2026, landing pages, brand voice, ads, cold email and UX copy, with honest caveats and install steps.

Ten curated agent skills for content writing in 2026, social posts, technical docs, UX strings and academic writing, with honest caveats and install steps.

Eight agent skills that push AI-generated design away from generic templates, Impeccable, Taste, Anthropic's Frontend Design and more, with honest caveats on each.

Ten curated agent skills for UI/UX work in 2026, design systems, wireframing, theming and anti-AI-slop critique, with honest caveats and install steps.

A curated selection of the strongest data analysis skills for AI agents, spreadsheet cleanup, Power BI modeling, SQL reconciliation and more, with honest caveats.

Ten agent skills worth installing for application security and code scanning in 2026 (vulnerability review, secret scanning, threat modeling and supply-chain checks) picked for quality and upkeep.

OpenAI's own models now read the same SKILL.md format Anthropic published. Here's the 2026 model timeline, what a 1M-token context window actually changes for skills, and how Codex CLI handles them.

Ten agent skills worth installing for DevOps and deployment work in 2026 (monorepo builds, dependency automation, dev environments and Azure infrastructure) picked for quality and upkeep.

Ten curated agent skills for documentation in 2026, codebase onboarding docs, ADRs, changelogs and coding standards, with honest caveats and install steps.

Nine agent skills worth installing for code review in 2026 (quality audits, duplication checks, PR triage and release review) picked for quality, popularity and upkeep.

Nine agent skills worth installing for debugging and testing in 2026 (trace inspection, runtime investigation, root-cause workflows and log analysis) picked for quality and upkeep.

Ten agent skills worth installing for frontend work in 2026 (forms, component testing, visual QA, animation and CSS guidance) picked for quality, popularity and upkeep.

How to install Agent Skills in Cline, the open-source autonomous coding agent for VS Code, via .cline/skills, CLI install, manual setup, and VS Code-specific troubleshooting.

How to install Agent Skills in OpenCode, SST's open-source terminal agent, via the .opencode/skills directory, the CLI command, manual setup, and troubleshooting.

How to ship Agent Skills to production Claude API agents through container uploads and the code execution tool, no CLI install command exists for this platform.

Both read the same SKILL.md standard, but activation, timing, and install paths differ. Here's how Claude Code and Codex CLI compare for running skills.

How to install Agent Skills in GitHub Copilot's agent mode using the .github/skills folder, the CLI command, manual setup, verification, scope, and fixes for common failures.

What to check before installing an agent skill: bundled scripts, permissions, prompt injection risk, and a concrete pre-install checklist.

How to install Agent Skills in Windsurf's Cascade agent: the skills CLI, manual setup in .windsurf/skills, personal vs project scope, and troubleshooting.

AGENTS.md loads on every request and should stay short. Skills load conditionally and can be as detailed as they need. Here's the practical split, worked out.

Subagents buy isolation and parallelism at a cost. Skills buy expertise cheaply in the current context. Here's how to tell which problem you actually have.

How to install Agent Skills in Google Antigravity: the skills CLI, manual setup in .agents/skills, the older .agent/skills path, scope, and troubleshooting.

MCP connects agents to live external systems over a protocol. Agent Skills are markdown instructions loaded from disk. Here's when to use each, and how they combine.

How to install Agent Skills in Cursor 2.4 and later: the skills CLI, manual setup in ~/.cursor/skills, personal vs project scope, and fixing skills that won't trigger.

Discovery, activation, execution: the three-phase pipeline behind every agent skill, and why the description field is the entire discovery surface.

How to install Agent Skills in OpenAI Codex CLI, invoke them directly with $skill-name, choose personal vs project scope, and fix skills that won't load.

A step-by-step guide to writing your own agent skill, from folder structure to the description field, with a full worked example.

How to install Agent Skills in the Claude desktop and web apps through Settings, without a terminal, upload steps, verification, scope limits, and troubleshooting.

A field-by-field breakdown of the SKILL.md format: frontmatter, body structure, and the scripts, references and assets folders a skill can bundle.

A complete walkthrough for installing Agent Skills in Claude Code: the skills CLI, manual installs, personal vs project scope, verification, and troubleshooting.

Agent Skills are portable folders of instructions that let AI agents like Claude Code and Cursor load expertise on demand. Here is how the format works, end to end.

How to install and download Claude skills on any agent, the skills CLI, manual folder copying, GitHub installs, and the exact path for Claude Code, Claude Desktop, and 14 more platforms.