Cursor SDK Custom Tools vs Claude Agent SDK vs OpenAI Agents SDK (June 2026)
Cursor SDK Custom Tools vs Claude Agent SDK vs OpenAI Agents SDK (June 2026)
The SDKs for building programmable AI coding agents matured in mid-2026. Cursor’s June 2026 release added custom tools and stores, putting it in direct competition with Claude Agent SDK and OpenAI Agents SDK. The three SDKs now overlap meaningfully in capability but remain opinionated toward different deployment surfaces. Here’s a practical comparison for picking the right one for your next project.
Last verified: June 17, 2026.
TL;DR
- Cursor SDK (June 2026 release): IDE-embedded agents, custom tools, custom stores, auto-review flows.
- Claude Agent SDK (mature, June 15 billing change): Standalone agents, sub-agents, skills, dynamic workflows.
- OpenAI Agents SDK (most mature): Production-grade, portable, GPT and Operator tooling.
- Pick by surface: IDE → Cursor. Standalone Claude → Claude Agent SDK. Anything else → OpenAI Agents SDK.
- All three are real platforms. None is clearly best for all use cases.
What’s in each SDK
Cursor SDK (June 2026)
Released as part of the June 2026 Cursor update, the SDK turns Cursor into a programmable agent platform inside the IDE.
| Feature | Notes |
|---|---|
| Custom Tools | Expose TypeScript or Python functions to the Cursor agent |
| Custom Stores | Pluggable state and memory backends |
| Auto-Review Flows | Programmable Bugbot + your validators |
| Bugbot improvements | 3x faster, 22% cheaper, 10% more bugs found |
| Design Mode (3.7) | Voice input, multi-select |
| Model routing | Choose Claude, GPT, Gemini, or custom |
Best fit: extending Cursor for your dev team. Worst fit: running an agent your customers use directly.
Claude Agent SDK
Anthropic’s standalone-agent SDK. Mature, opinionated, Claude-first. The June 15, 2026 billing change moved Agent SDK and claude -p headless usage to separate API-billed credits (no longer covered by Pro / Max subscription pools).
| Feature | Notes |
|---|---|
| Sub-agents | First-class primitive, well-tuned |
| Skills | Reusable agent capabilities, share across projects |
| Dynamic workflows | Agents that orchestrate other agents |
| Tool calling | Type-safe, well-documented |
| Model routing | Claude Opus 4.7, Sonnet 4.5, Haiku 4.5 (Fable 5 and Mythos 5 currently suspended per June 12 export-control order) |
| Billing model | API credits as of June 15, 2026 |
Best fit: production Claude-powered agents with sub-agent / skill composition. Worst fit: multi-provider agents or maximum portability.
OpenAI Agents SDK
The longest-running of the three. Python and TypeScript SDKs, GA since mid-2025, with the broadest production track record.
| Feature | Notes |
|---|---|
| Production maturity | Most-used in production today |
| Provider portability | Easy to swap models inside the SDK |
| Operator integration | Built-in agentic browsing capability |
| Code Interpreter | First-class sandbox execution |
| File Search | Built-in document RAG |
| Model routing | GPT-5.5, plus support for non-OpenAI models |
| Billing model | Per-token API billing |
Best fit: agents that need to run anywhere, use OpenAI’s tool ecosystem, or interop with multiple model providers. Worst fit: tightly Claude-bound use cases where Claude Agent SDK’s sub-agent ergonomics are better.
Capability head-to-head
| Capability | Cursor SDK | Claude Agent SDK | OpenAI Agents SDK |
|---|---|---|---|
| Custom tool calls | ✅ | ✅ | ✅ |
| Custom state / memory | ✅ Stores | ✅ Skills + workflows | ✅ |
| Sub-agents | Limited | ✅ Mature | ✅ Via orchestration |
| Multi-provider routing | ✅ | Anthropic-first | ✅ Easy |
| IDE integration | ✅ Native | Via plugins | Via plugins |
| Standalone runtime | ❌ | ✅ | ✅ |
| Production maturity | New | Mature | Most mature |
| Voice / multimodal | ✅ Design Mode | Via Claude models | ✅ Realtime API |
| Sandbox execution | Limited | Via skills | ✅ Code Interpreter |
| Browser agent | Limited | Limited | ✅ Operator |
| Documentation depth | Growing | Strong | Strongest |
When each one wins
Choose Cursor SDK when:
- You want your dev team to have agent-powered code review inside Cursor
- You’re building internal developer tooling that fits naturally in an IDE
- You value the IDE-native UX over running the agent elsewhere
- You want Bugbot extended with your own validators
Choose Claude Agent SDK when:
- You’re building a standalone Claude-powered agent
- Sub-agent composition matters (research → analysis → write)
- You want skills as a reusable abstraction
- Your team is heavily Anthropic-aligned and OK with single-vendor risk
- You can absorb the June 15 billing change (API credits vs Pro/Max pool)
Choose OpenAI Agents SDK when:
- Production maturity is the primary criterion
- You need provider portability (model swaps without rewriting)
- You’re already in the OpenAI ecosystem (Operator, Code Interpreter)
- You want the broadest community and the most production examples
- You’re building an agent that ships as a product
What about LangGraph, Mastra, and the other agent frameworks?
LangGraph, Mastra, and Autogen remain relevant but increasingly orchestrate the SDK-level work rather than competing with it. The common 2026 pattern: use LangGraph or Mastra as your workflow orchestration layer, then call into Claude Agent SDK or OpenAI Agents SDK for the actual agent execution. Cursor SDK doesn’t compete with this layer at all — it’s a different surface (IDE), not an orchestration framework.
If you’re picking an orchestration framework: LangGraph for production Python shops, Mastra for TypeScript shops, plain SDK calls for simple agents that don’t need heavy workflow logic.
Pricing reality check (June 2026)
| SDK | Effective pricing |
|---|---|
| Cursor SDK | Bundled with Cursor subscription ($20-40/mo Business+) |
| Claude Agent SDK | API rates: ~$15/$75 per MTok for Opus 4.7 |
| OpenAI Agents SDK | API rates: ~$5/$15 per MTok for GPT-5.5 standard |
For sustained programmatic use, OpenAI Agents SDK on GPT-5.5 is currently the cheapest per quality unit. Claude Agent SDK on Opus 4.7 is the highest quality on coding-specific tasks but costs 3-4x more per token. Cursor SDK is the cheapest if you’re already paying for Cursor anyway and the agent runs inside the editor.
What changes when Fable 5 comes back
If the Anthropic export-control situation resolves and Fable 5 returns to availability, the Claude Agent SDK story improves significantly — Fable 5 was the SWE-bench Pro leader pre-suspension. Production agents that need top-tier coding capability would naturally upgrade. Pricing for Fable 5 was at the high end ($10/$50 per MTok structure was floated), so the OpenAI Agents SDK cost advantage on standard tasks remains. Cursor SDK isn’t directly affected — it routes through whichever model you pick.
Bottom line
Three real SDKs, three different surfaces. Most production agents end up in OpenAI Agents SDK because portability and maturity matter most. Claude-first shops should use Claude Agent SDK for the better sub-agent ergonomics. Cursor SDK is the right choice for IDE-embedded developer tooling, not for standalone agents.
Pick by your deployment topology and the model you’re committed to. Migration between SDKs is annoying but doable — don’t over-think the choice, ship the agent, and revisit in 6 months.