What is the Cursor SDK and what does it do?

The Cursor SDK is a programmatic agent SDK Cursor announced on April 29, 2026 and elevated in their May 4, 2026 'Build programmatic agents with the Cursor SDK' update. It lets developers run Cursor's coding agent (powered by Composer 2, GPT-5.4/5.5, Opus 4.6, Gemini 3 Pro, or Grok Code) headlessly inside their own scripts, CI pipelines, or applications. Pattern: import { Agent } from '@cursor/sdk', create an agent with a model and working directory, send a task, stream events. It's Cursor's answer to Claude Code's CLI-as-SDK and OpenAI's Codex CLI.

Which agent SDK should I pick for a custom AI coding integration?

Three rules. (1) You're building inside Cursor IDE workflows or want IDE-quality codebase indexing in your headless flow → Cursor SDK — it inherits Cursor's indexing and agent orchestration. (2) You want best-in-class agentic stamina and tightest Anthropic integration (Skills, MCP, Claude Code patterns) → Claude Agent SDK. (3) You're OpenAI-native or want first-class GPT-5.5 + Codex parity → OpenAI Agents SDK. The SDKs aren't substitutes — they each pull in their parent product's strengths and lock-in.

Are these SDKs free or do they cost money?

All three SDKs are free to install. Inference costs vary. Cursor SDK requires a Cursor API key and bills against Cursor pricing (Pro $20/mo or usage-based) plus the chosen model's costs. Claude Agent SDK uses Anthropic API tokens (pay-per-token) or Claude Code subscription. OpenAI Agents SDK uses OpenAI API tokens. For high-volume programmatic agent work, OpenAI is typically cheapest per-token at GPT-5.4 prices, Anthropic is most expensive per-token but most capable per task on Opus 4.6/4.7, Cursor sits in the middle but bundles indexing services.

Can I use multiple SDKs together?

Yes, and it's becoming common. Pattern: use one SDK as the orchestrator and call others as tools. For example, Cursor SDK can spawn a Claude Code agent for a long Opus 4.7 task, or OpenAI Agents SDK can use Cursor SDK for codebase-aware retrieval. All three SDKs support MCP servers, so they can share tooling (AWS MCP Server, GitHub MCP, etc.). Multi-SDK setups give you model and capability optionality without re-platforming the whole agent flow.

Quick Answer

Cursor SDK vs Claude Agent SDK vs OpenAI Agents SDK (May 2026)

Published: May 7, 2026

Cursor SDK vs Claude Agent SDK vs OpenAI Agents SDK (May 2026)

On May 4, 2026, Cursor elevated its programmatic agent SDK in a ‘Build programmatic agents with the Cursor SDK’ release — turning Cursor’s coding agent into a headless, scriptable building block. It joins Anthropic’s Claude Agent SDK and OpenAI’s Agents SDK as the three main coding-agent SDKs in May 2026. Here’s how they compare.

Last verified: May 7, 2026

The three at a glance

Capability	Cursor SDK	Claude Agent SDK	OpenAI Agents SDK
Vendor	Cursor (Anysphere)	Anthropic	OpenAI
Released	April 29, 2026 (programmatic agents)	Late 2025, expanded 2026	2025 (Agents SDK), Codex SDK 2026
Model lineup	Composer 2, GPT-5.4/5.5, Opus 4.6, Gemini 3 Pro, Grok Code	Anthropic only (Sonnet 4.6, Opus 4.6/4.7, Mythos preview)	OpenAI only (GPT-5.4, GPT-5.5, GPT-5.3-Codex)
Codebase awareness	Cursor’s full indexing	Skills + MCP	OpenAI Codex tools
MCP support	Yes	Yes (native)	Yes
Headless / scriptable	Yes	Yes	Yes
Streaming events	Yes	Yes	Yes
Best for	Cursor-flavored agents in apps / CI	Long-horizon Anthropic-stack agents	OpenAI-native pipelines

What “agent SDK” means in May 2026

All three SDKs share a common shape:

Create an agent instance with a model selection, working directory, and tools / MCP servers.
Send a task (natural language).
Stream events as the agent reads files, runs commands, calls tools, and edits code.
Receive a final result (success / failure / artifacts).

What differs is whose product the SDK pulls in. Each SDK is the headless mode of its parent agent — Cursor SDK is Cursor minus the IDE, Claude Agent SDK is Claude Code minus the CLI, OpenAI Agents SDK is Codex minus its CLI/desktop wrapper.

Cursor SDK

import { Agent } from "@cursor/sdk";
const agent = await Agent.create({
  apiKey: process.env.CURSOR_API_KEY!,
  model: { id: "composer-2" },
  local: { cwd: process.cwd() },
});
const run = await agent.send("Summarize what this repository does");
for await (const event of run.stream()) {
  console.log(event);
}

What you get:

Cursor’s codebase indexing — same one that powers Cursor IDE — applied to any directory.
Full model choice across Composer 2 (Cursor’s own coding model), GPT-5.4, GPT-5.5, Opus 4.6, Gemini 3 Pro, Grok Code.
Streaming agent events (file reads, tool calls, code edits).
Local-first execution with optional cloud sandbox.
MCP server support.

Strength: the codebase indexing is the best-in-class component of Cursor and now usable headlessly.

Claude Agent SDK

Anthropic’s SDK (also exposed as claude-code CLI flags like --print) lets you script Claude Code:

Best-in-class Opus 4.6/4.7 agentic stamina on long tasks.
Native MCP support — drops in any MCP server (AWS MCP, GitHub MCP, Atlassian, etc.).
Skills system — declarative agent capabilities.
Headless CLI mode (claude --print --permission-mode bypassPermissions) that works as a drop-in SDK in any language via shell.
Claude Code Cloud variant for managed runs without local execution.

Strength: highest-quality model for long-horizon agentic work; tight integration with the broader Anthropic agentic stack.

OpenAI Agents SDK

OpenAI’s Agents SDK (Python and TypeScript) plus the Codex CLI provide the OpenAI-flavored equivalent:

GPT-5.4, GPT-5.5, GPT-5.3-Codex (the combined coding+reasoning model).
Tight integration with OpenAI’s tool ecosystem (Code Interpreter, function calling, MCP).
Codex CLI parity — same agent loop the Codex CLI uses.
Now available on Amazon Bedrock as of the May 2026 launch (Bedrock Managed Agents powered by OpenAI).
Best-in-class for parallel execution and structured outputs.

Strength: cheapest per-token of the three on GPT-5.4 / 5.3-Codex; best AWS-native story after the Bedrock launch.

Where each one wins

Cursor SDK wins for…

Custom apps that want Cursor-quality codebase indexing programmatically.
CI / pre-commit / pre-PR automation that benefits from indexed retrieval.
Multi-model workflows — Cursor SDK is the only one with first-class support across all major frontier models.
Teams already in the Cursor ecosystem who want one SDK across IDE + automation.
Use cases where you want an agent that “feels” like Cursor’s IDE agent.

Claude Agent SDK wins for…

Long-horizon multi-file agentic tasks where Opus 4.6/4.7 stamina matters.
Workflows that lean on Anthropic Skills and the broader MCP ecosystem.
Customers already standardized on Anthropic.
Use cases needing the best model quality regardless of cost.
Agents that hand off between Claude chat / Claude Code / programmatic via shared MCP.

OpenAI Agents SDK wins for…

OpenAI-native shops with GPT-5.5 / GPT-5.3-Codex.
High-volume agent workloads where token cost matters.
AWS-native enterprises after the Bedrock Managed Agents launch.
Workflows needing parallel execution at scale.
Pipelines integrating Code Interpreter for data work.

Multi-SDK patterns

Real production deployments in May 2026 are increasingly multi-SDK:

Pattern: orchestrator + specialist. One SDK orchestrates; another handles specific subtasks. Example: Cursor SDK orchestrates a refactor across 50 files; spawns Claude Agent SDK runs for the 5 hardest files needing Opus 4.7 reasoning.
Pattern: cost tiering. Use OpenAI Agents SDK with GPT-5.4 for cheap routine work; route to Claude Agent SDK with Opus 4.7 only when GPT-5.4 fails. This is the “AI Router” pattern.
Pattern: shared MCP. All three SDKs hit the same MCP servers (AWS MCP Server, GitHub MCP). Tooling stays shared; models / agents differ.

The MCP layer is what makes multi-SDK flows clean — all three SDKs are MCP clients, so tools written once work across all three.

Pricing

SDK	SDK cost	Inference cost
Cursor SDK	Free	Cursor API ($20/mo Pro tier or usage-based) + model token cost
Claude Agent SDK	Free	Anthropic API tokens (or Claude Code subscription)
OpenAI Agents SDK	Free	OpenAI API tokens (or via Bedrock pricing)

For typical agentic coding work in May 2026:

Cheapest per-token: OpenAI GPT-5.4.
Most expensive per-token: Claude Opus 4.7 / Mythos preview.
Best mid-cost option: Cursor with Composer 2 (Cursor’s own model, optimized for cost).

Cost varies wildly by workload. For lightweight completion, OpenAI wins; for hardest agentic tasks, Anthropic often costs less per task because it succeeds first try.

The MCP question

All three SDKs support MCP, but with subtle differences:

Cursor SDK supports MCP servers configured per-project.
Claude Agent SDK supports MCP natively at the protocol level (Anthropic created MCP).
OpenAI Agents SDK added first-class MCP support in early 2026; now treats MCP as a primary tool source.

In practice, MCP works equivalently in all three — write a tool once as an MCP server, use it from any SDK. This is the design Anthropic and Linux Foundation Agentic AI Foundation pushed; it’s working.

Bottom line

In May 2026, the three coding-agent SDKs are now genuinely competitive — pick the one that matches your existing stack and model preferences, not based on raw capability differences. Cursor SDK wins on multi-model flexibility and best codebase indexing. Claude Agent SDK wins on agentic stamina and Anthropic-stack integration. OpenAI Agents SDK wins on cost at scale and OpenAI / AWS-native pipelines. Most serious agentic coding deployments will use two or three together, glued by MCP — let the SDK question be a tactical pick per workflow, not a strategic one.

Sources: Cursor changelog (May 4, 2026), Cursor SDK launch April 29, 2026, Releasebot Cursor / OpenAI updates (May 2026), Anthropic Claude Code documentation (May 2026), OpenAI Codex changelog (May 2026), AWS Bedrock Managed Agents launch (May 2026).