Anthropic Agent SDK adoption surges — native permissions & sessions hooks in 100+ repos

Anthropic’s Claude Agent SDK is spreading quickly through GitHub and npm as developers package Claude Code’s agent loop into apps with built-in permissions, hooks, sessions and subagents. (code.claude.com) (npmjs.com) The SDK is the renamed Claude Code SDK, and Anthropic says it gives developers the same tools, context management and agent loop that power Claude Code itself in Python and TypeScript. (code.claude.com) (github.com 1) (github.com 2) Anthropic’s TypeScript package was published at version 0.2.119 in the last day and lists 974 npm dependents, while the Python repo shows 6.5 thousand stars and 923 forks as of April 26, 2026. (npmjs.com) (github.com) The technical pitch is control. Anthropic’s permissions docs say tool calls can be filtered through hooks, deny rules, permission modes, allow rules and a `canUseTool` callback before an agent gets to act. (code.claude.com) That matters for developers trying to let an agent read files, edit code or run shell commands without handing it unlimited access. Anthropic says deny rules still override even in `bypassPermissions` mode, and project rules can live in `.claude/settings.json`. (code.claude.com) Subagents are the other big draw. Anthropic says each subagent runs in its own context window with its own system prompt, tool access and permissions, then returns only a summary to the parent session. (code.claude.com) Anthropic’s own demo repo now includes a research agent that breaks a topic into subtopics, spawns parallel researcher agents and synthesizes the results, plus a Hello World V2 example for multi-turn session persistence. (github.com) Outside Anthropic, developers are building bigger scaffolding around the SDK. AnthroClaw, a repo updated within the last day, describes itself as a “Claude Agent SDK-native control plane” for Telegram, WhatsApp and web chat with sessions, metrics, checkpoints, hooks, skills and MCP tools. (github.com) The surrounding ecosystem is also getting large enough to look like a marketplace. VoltAgent’s `awesome-claude-code-subagents` repo says it offers 100-plus specialized subagents, and wshobson’s `agents` repo advertises 184 specialized agents, 150 skills and 16 workflow orchestrators. (github.com 1) (github.com 2) Other repos are turning Claude setups into reusable project assets. Feiskyer’s `claude-code-settings` repo packages settings, hooks, plugins, skills and subagents, while Anthropic’s docs point developers to `.claude` files such as `CLAUDE.md`, hooks and settings as the place where that behavior lives. (github.com) (code.claude.com) Anthropic is also drawing a line around how the SDK should be used. Its docs say developers should authenticate with API keys or supported cloud providers such as Amazon Bedrock, Google Vertex AI and Microsoft Azure AI Foundry, and warn that third parties are not allowed to offer Claude.ai logins or Anthropic rate limits for SDK-based products. (code.claude.com) The result is a fast-growing stack of official SDKs, demos and community repos that treat Claude less like a chat box and more like an operating layer for software agents. (code.claude.com) (github.com)