# Cursor SDK vs Claude Agent SDK: languages, models and billing

> Claude's SDK runs Claude models in your own process. Cursor's runs on its hosted models and can deploy agents to its cloud. Here is how they differ.

URL: https://claudegpt.si/news/cursor-sdk-vs-claude-agent-sdk  
Published: 2026-10-29T02:51:09.269Z  
Author: ClaudeGPT Si Newsroom  
Section: Reviews  
Topics: Cursor, Claude Code, Anthropic, AI agents, Developer tools, Comparison  
Publisher: ClaudeGPT Si

## Key takeaways

- The Claude Agent SDK runs Claude Code as a library in Python or TypeScript, on Claude models, billed as Anthropic token cost.
- Cursor's @cursor/sdk and cursor-sdk call Cursor's agent on Cursor's hosted models, billed to your Cursor plan or team.
- Cursor's Agent SDK, announced October 20, 2026, adds deployable agents, but its docs state no pricing or full model list.

Cursor's SDK and the Claude Agent SDK both let you put a coding agent inside your own software, but they differ on three points that matter. Claude's SDK is a library for Python and TypeScript that runs on Claude models inside a process you operate. Cursor's SDK has TypeScript and Python versions, runs on Cursor's hosted models, and its separate Agent SDK can deploy agents to Cursor's cloud. This cursor sdk vs claude agent sdk comparison is based on the products' docs and published results, as read on October 29, 2026, not on hands-on testing.

"Cursor SDK" covers two things in Cursor's docs. [`@cursor/sdk`](https://cursor.com/docs/sdk/typescript) lets you script the same agent that runs in the Cursor editor and CLI. The [Agent SDK](https://cursor.com/blog/agent-sdk), which Cursor announced on October 20, 2026, is a framework for defining and deploying agents as code. We cover both, since people mean either. The Anthropic side is the [Claude Agent SDK](https://code.claude.com/docs/en/agent-sdk/overview), which Anthropic describes as Claude Code as a library.

## Cursor SDK vs Claude Agent SDK at a glance

| | Claude Agent SDK | Cursor `@cursor/sdk` | Cursor Agent SDK |
| --- | --- | --- | --- |
| Languages | Python and TypeScript | TypeScript, and Python as `cursor-sdk` | TypeScript and Markdown |
| What it runs | A bundled Claude Code binary, started as a subprocess of your app | Cursor's agent, with the loop local or in a Cursor-hosted VM | A project of prompts, tools and evals, run locally, on Cursor's hosting or self-hosted |
| Models | Claude models, chosen by alias or full name, with an optional fallback model | Cursor's hosted models, such as `composer-2.5`, plus a Router on Teams and Enterprise | Docs show `composer-2.5` and default to `grok-4.5`; no full list given |
| Billing | Anthropic API token costs | Your Cursor plan or team, same rules as the IDE | Not stated in the docs we read |
| Auth | `ANTHROPIC_API_KEY`, or Bedrock, Vertex, Foundry | `CURSOR_API_KEY` or browser login | `agent-sdk login`; `CURSOR_API_KEY` when self-hosting |
| Runtime needed | Node.js 18.14+ or Python 3.10+ | Node.js 22.13+ | Node.js 22.13+ |

## Languages and what each one runs

The Claude Agent SDK ships as `@anthropic-ai/claude-agent-sdk` for TypeScript and `claude-agent-sdk` for Python. Both bundle a native Claude Code binary. Per Anthropic's [hosting docs](https://code.claude.com/docs/en/agent-sdk/hosting), your `query()` call starts a separate `claude` process and talks to it over stdio. Each session is one long-lived process with its own shell, working directory and session files. To use another language, Anthropic says to run the CLI as a subprocess with the `-p` flag.

Cursor's `@cursor/sdk` is a TypeScript package, and a Python package, `cursor-sdk` for Python 3.10 or later, offers the same agent. You pick `local` or `cloud` when you create an agent. Local means the loop runs in your Node process and reads files from disk. Cloud means the agent runs in an isolated VM with your repo cloned in, and runs can continue after the caller disconnects. In both cases, inference runs on Cursor's hosted models.

The Cursor Agent SDK works differently. An agent is a folder of Markdown prompts and TypeScript tools, configured with `defineAgent`, that you test with evals and deploy with `agent-sdk deploy`. Cursor says it uses this to build Bugbot, Security Reviewer and Rollouts. You can host the result on Cursor or run `agent-sdk serve` yourself under Node 22.13 or newer.

One oddity: the announcement names the package `@cursor/agent-sdk`, while the docs install `@grok/agent-sdk` and say it replaces `@cursor/bdk`. Check the current name before you install.

## Models supported

The Claude Agent SDK runs Claude models. You set `model` to an alias or full model name, and `fallbackModel` names a backup that takes over when the primary is overloaded. Anthropic's docs also list the SDK's support for Amazon Bedrock, Google Cloud's Agent Platform and Microsoft Foundry, which are still Claude models served through those clouds.

Cursor's docs show a narrower, Cursor-managed catalog. `Cursor.models.list()` returns the models available to your account, and the docs recommend caching it instead of hard-coding ids. The `@cursor/sdk` page names `composer-2.5`, with parameters such as `fast` and reasoning effort. Teams and Enterprise accounts can use Cursor Router, which picks a model per request and takes an `optimize_for` setting of `cost`, `balanced` or `intelligence`. The Agent SDK reference defaults to `grok-4.5` and does not list the other models or providers.

## Pricing and billing

Pricing is where the two differ most, and where the docs leave the most gaps. As of October 29, 2026:

- **Claude Agent SDK:** the docs describe usage as Anthropic token cost. The SDK reports `total_cost_usd`, but Anthropic calls it a client-side estimate, not billing data. Authoritative numbers come from the Usage and Cost API or the Claude Console. Anthropic's hosting page says a minimal container costs about $0.05 an hour, and that token cost usually dwarfs it.
- **Cursor `@cursor/sdk`:** SDK runs follow the same pricing, request pools and Privacy Mode rules as IDE and Cloud Agent runs. Service account keys bill to the owning team, and user keys bill to the user's plan. Spend shows up under an SDK tag in the team usage dashboard. The page gives no per-token rates. For plan prices, see our [Cursor pricing](https://claudegpt.si/news/cursor-pricing-plans-weekly-pool) article.
- **Cursor Agent SDK:** neither the announcement nor the docs pages we read state pricing or usage limits. The only numbers are deployment limits: 20 deployments per team and 100 deploys per rolling 24 hours, if your team has caps.

## Auth

Claude's SDK reads `ANTHROPIC_API_KEY` from the environment of the process that runs your agent. It also supports Bedrock, Google Cloud's Agent Platform and Microsoft Foundry through environment variables. Anthropic adds that, unless previously approved, third-party products built on the SDK may not offer claude.ai login or its rate limits, so products should use API keys.

Cursor's `@cursor/sdk` takes a user API key or a service account API key through `CURSOR_API_KEY`. Team Admin keys are not yet supported. `Cursor.auth.login()` opens a browser login and mints a user key that lasts 90 days by default. The Agent SDK uses `agent-sdk login` for model turns, `CURSOR_API_KEY` for self-hosted runs, and a bearer token on a self-hosted server. Secrets are declared by name in code and set with `agent-sdk secret set`.

## Features

Both cover the basics of a coding agent: streaming events, subagents, MCP servers, sessions you can resume, and hooks. The differences are in what surrounds them.

- **Claude Agent SDK:** built-in tools that read and edit files, run commands and search the web, plus permission modes, skills, plugins, structured output, file checkpointing, spend and turn caps (`maxBudgetUsd`, `maxTurns`) and OpenTelemetry export. Hooks can be code callbacks.
- **Cursor `@cursor/sdk`:** local and cloud runtimes, `run.steer()` to redirect a running local turn, optional sandboxing, cloud artifacts and per-run usage through `getUsage()`. Its hooks are file-based only, with no programmatic callbacks.
- **Cursor Agent SDK:** event triggers from pull requests, Slack and issue trackers, evals with event replay, a local playground, and a dashboard for deployed agents. It is aimed at agents that react to events in a team's workflow.

## Which one fits

Pick the Claude Agent SDK if you work in Python, want Claude models, or need to run the agent in your own infrastructure. Pick Cursor's SDKs if you already pay for Cursor and want its hosted models and deployment, with the caveat that the Agent SDK's pricing and model list are not documented yet. For the products these SDKs come from, see our [Cursor vs Claude Code comparison](https://claudegpt.si/news/cursor-vs-claude-code-comparison).

## Frequently asked questions

### Is the Cursor SDK the same as the Cursor Agent SDK?

No. @cursor/sdk calls Cursor's own agent from your code. The Agent SDK is a separate framework for defining agents as Markdown and TypeScript projects and deploying them. Cursor's docs treat them as different pages.

### Does the Claude Agent SDK work with models other than Claude?

Its docs describe Claude models only. It can reach them through Anthropic's API, Amazon Bedrock, Google Cloud's Agent Platform or Microsoft Foundry.

### Which agent SDK supports Python?

Both vendors offer a Python package: claude-agent-sdk for Python 3.10 or later, and Cursor's cursor-sdk, also for 3.10 or later. Cursor's Agent SDK framework is TypeScript and Markdown.

### Is the Claude Agent SDK free?

The docs list no SDK fee. Usage is billed as Anthropic token cost through your API key or cloud provider, and you pay for any infrastructure you run it on.

## Sources

- [Cursor Agent SDK announcement](https://cursor.com/blog/agent-sdk), Cursor
- [Cursor SDK (TypeScript) docs](https://cursor.com/docs/sdk/typescript), Cursor
- [Cursor SDK (Python) docs](https://cursor.com/docs/sdk/python), Cursor
- [Cursor Agent SDK docs](https://cursor.com/docs/agent-sdk), Cursor
- [Cursor Agent SDK deployment](https://cursor.com/docs/agent-sdk/deployment), Cursor
- [Agent SDK overview](https://code.claude.com/docs/en/agent-sdk/overview), Anthropic
- [Agent SDK quickstart](https://code.claude.com/docs/en/agent-sdk/quickstart), Anthropic
- [Configure your agent](https://code.claude.com/docs/en/agent-sdk/configuration), Anthropic
- [Hosting the Agent SDK](https://code.claude.com/docs/en/agent-sdk/hosting), Anthropic
- [Track cost and usage](https://code.claude.com/docs/en/agent-sdk/cost-tracking), Anthropic