The Claude Agent SDK is the Claude Code binary with an API around it
Anthropic's overview says the SDK gives you the same tools, agent loop and context management that power Claude Code, programmable in Python and TypeScript. Mechanically, it's a library that runs the Claude Code binary. Your process drives it. The Python package bundles the Claude Code CLI, according to Short Stack Daily.
The capability list is long. Built-in tools read, write and edit files, run commands and search the web. On top sit hooks, subagents, MCP servers, permissions, and sessions you can resume or fork. Skills, commands and memory load from your project's .claude/ folder, same as in the CLI. Plugins bundle all of it.
Anthropic draws its own map of neighbours. The CLI is for interactive work. The Client SDK is where you write the tool loop yourself. Managed Agents is the hosted option. The Agent SDK is for embedding the agent in a process you operate. For other languages, the docs suggest running the CLI as a subprocess with -p and --output-format json.
What you get for free, and why it matters on a Monday
The tool loop is the dull 80% of a coding agent. File edits, shell access, approval prompts and context handling are weeks of work if you build them on a raw model API. Here they exist and are the same ones people use daily in Claude Code. That's our opinion, but it's the main reason we'd start here for anything that touches a repository.
The permission layer is the part we'd read first. You decide which tools run automatically and which need approval, and hooks let you run your own code at key points in the lifecycle. According to the AI library digest, the latest Python and TypeScript SDK releases both added fail-closed hook matchers. We haven't read those changelog entries line by line. But fail-closed is the right default for the code that decides what an agent may do.
One more point in its favour: Infragentic's comparison lists OpenTelemetry export for it, so traces don't have to go to a vendor dashboard. That is their reading of the docs, not something we tested.
Where it bites: Claude only, 0.x versions, a binary underneath
Model choice is the first catch. Dbggr's comparison says the SDK runs only Claude, though through the Anthropic API, Bedrock, Google Cloud or Foundry. Pydantic AI, LangGraph, Mastra and Vercel's AI SDK switch models freely, in the same article's reading. If your cost plan depends on swapping models next quarter, that matters.
The second catch is churn. Infragentic counted Python 0.2.158 and TypeScript 0.3.280 on September 22 and described several releases a week. Dbggr saw Python 0.2.168 on October 15. Short Stack Daily saw 0.2.176 on October 31. Nothing here is 1.0. Pin your versions and read the changelog before every bump, because your agent's behaviour rides on a CLI binary that moves fast.
Licensing is the third. The overview says use is governed by Anthropic's Commercial Terms, except where a component carries its own licence. Sources disagree on the details: Short Stack Daily lists the Python SDK as MIT and sends you to the README for TypeScript, while Dbggr says only the Python wrapper is MIT and the TypeScript package and bundled CLI fall under the Commercial Terms. The October 21 move from MIT to Apache 2.0 that the digest reports covers Anthropic's new SDK majors (Python 2.0, TypeScript 1.0). The Agent SDK is still 0.x, so those are evidently the client SDKs. Read the LICENSE file in the exact package you ship.
Two rules from the docs also bind products. Third parties can't offer claude.ai login or its rate limits unless Anthropic approved it, so you use API keys. And your product can't be branded Claude Code or mimic its visuals.
Against LangGraph, the OpenAI Agents SDK and Managed Agents
Short Stack Daily's quick picks put LangGraph first for durable, stateful workflows with approvals, the OpenAI Agents SDK for lean agents with handoffs, and the Claude Agent SDK for coding and computer-work agents. We agree with the split. LangGraph has you draw the graph. The Claude SDK hands you a complete coding-agent setup, as Infragentic puts it. We covered the other two in our LangGraph teardown and our OpenAI Agents SDK teardown.
Durability is where the gap shows. The overview describes sessions that keep context and can be resumed or forked. It doesn't describe crash-proof durable execution of the kind the LangGraph docs advertise. If a run must survive a dead host, plan that part yourself.
If you don't want to run the agent at all, Anthropic sells Managed Agents. Dbggr prices it at $0.08 per running session-hour plus tokens, sandbox included, and says neither Anthropic's nor OpenAI's managed loop is eligible for zero data retention. The same article says no independent evidence shows one multi-agent design beating another, and that its author didn't run any of the frameworks. So treat subagents as a tool, not a virtue. For hedging, the digest notes an experimental agent type in Vercel's AI SDK 7 that wraps Claude Code, Codex and other coding agents behind one interface.
Who should build on it, and who should wait
Build on it if the agent's job is to work with files and a shell: repo maintenance, migration scripts, review bots, internal tools that fix or generate code. You inherit a tool loop that a lot of people use every day.
Wait if you need several model vendors, resumable long runs, or fixed step-by-step control. Also wait if you want customers to sign in with their own Claude accounts. For plain business flows (a webhook, an LLM call, a CRM update) a full agent loop is overkill, and we'd still reach for n8n or Make with one model call inside.
Strengths, weaknesses, where it fits
Strengths
- Built-in file, shell and web tools you don't have to write
- Permission controls and hooks for gating what the agent may do
- Skills, commands and memory load from your project's .claude folder
- Subagents and MCP servers supported out of the box
Weaknesses
- Runs only Claude models, via Anthropic's API or three clouds
- Still 0.x with several releases a week, so versions need pinning
- Licence terms vary by package and sources disagree on the details
- No documented durable execution; sessions resume but runs aren't crash-proof
Where it fits
- Repo maintenance and migration agents
- Code review and bug-fixing bots
- Internal tools that edit files and run commands
- Agents that should behave like Claude Code inside your own process
Verdict: We'd put it in production for coding and file-and-shell agents on Claude, behind a thin wrapper and with pinned versions. We wouldn't pick it for multi-vendor setups or long runs that must survive a crash.
Sources
- Agent SDK overview - Claude Code Docs · Anthropic
- Agent and LLM framework landscape · AI library
- Best AI Agent Frameworks (November 2026) · Short Stack Daily
- AI agent SDKs compared: OpenAI, Claude, LangGraph, Pydantic AI (2026) · Dbggr
- AI Agent Frameworks: 10 Compared and Verified for 2026 · Infragentic
Want this working in your stack, not just in a demo?
We are two senior engineers who build ai workflow automation with a fixed price per written scope.