Compare

A short cut of the comparison table from the main README.

2026-06-17 update. The framework field has shifted: OpenAI Agents SDK now has a native sandbox (2026-04), Vercel AI SDK 6 ships DurableAgent + a DevTools panel, Cloudflare ships code-mode portal-default. Several rows that used to be wasmagent-only are now table stakes. The differentiation has tightened to swappability (drop our kernel into any framework's executor socket) and governance (CapabilityManifest ↔ OWASP Agentic Top 10 with real WASM isolation; see docs/security/capability-manifest-owasp.md).

	Vercel AI SDK 6	Mastra	LangGraph.js	OpenAI Agents JS	CF Agents SDK	wasmagent-js
Edge-safe sandboxed code execution	❌	❌	❌	✅ native sandbox (2026-04)	✅ isolate (CF only)	✅ 3 tiers, portable
Real Python in-process	❌	❌	❌	❌	❌	✅ Pyodide
Auto prompt-cache breakpoints + 1h TTL	❌	❌	❌	❌	❌	✅
Code-mode MCP server (docs_search + execute_code)	❌	❌	❌	⚠️ via Agents SDK builtin	✅ portal-default (2026-03)	✅
Chinese models (Doubao / DeepSeek / Kimi / Qwen / GLM / MiniMax)	❌	❌	❌	❌	❌	✅
Time-travel debugger + fork-from-step	❌	✅ Studio	✅ Studio	❌	❌	✅
Stateless HITL + SSE Last-Event-ID resume	✅ DurableAgent (2026-Q1)	⚠️ partial	✅	❌	⚠️ DO only	✅
Local Studio (cost / latency / errors)	✅ DevTools panel	✅ Studio (SaaS)	❌	❌	❌	✅ zero-deploy, no telemetry
Pareto-first multi-model evaluation harness	❌	❌	❌	❌	❌	✅
Reproducible benchmarks gated by CI	❌	❌	❌	❌	❌	✅
Framework-neutral CapabilityManifest + WASM isolation (OWASP-mapped)	❌	❌	❌	⚠️ sandbox without policy primitive	⚠️ isolate without portable policy	✅
Drop-in executor for other frameworks' sandbox sockets	(target: AI SDK provider)	✅ mastra-sandbox	n/a	(target: Agents SDK provider)	(target: codemode DynamicWorkerExecutor)	✅

"❌" means the framework does not ship the capability today (2026-06). Where alternatives are partially available, "⚠️" notes the caveat. We re-verify quarterly. The two new rows at the bottom were added 2026-06-17 to reflect the tightened differentiation.

Where we are not

We are intentionally not competing on:

• Number of integrations — LangChain has 300+; we won't catch up on raw breadth.
• React-first DX — Vercel AI SDK 6 ships as the default in Next.js templates; we don't try to dethrone that.
• Hosted commercial SaaS — Mastra Cloud, LangSmith, Vercel platform. We're a runtime + framework, not a hosted product.
• Headline benchmark numbers. Agent benchmarks are in a "self-reported, never independently reproduced" trust crisis (2026-Q1/Q2 industry consensus). Our angle is the referee, not the contestant — evals-runner (paired McNemar / Wilson / bootstrap, scipy-aligned 1e-7) is the harness anyone can re-run on any pair of agents. Worked example: three-round arm-f vs bare vs batch-grammar ablation.

If you're choosing between wasmagent-js and one of the above frameworks, the question is whether you need any of the rows in the table that say "✅" only for wasmagent-js. If you don't, the bigger ecosystems are the right call.

Why three sandbox tiers?

See the kernel decision tree — every cell other than wasmagent-js is "❌" because no framework treats sandbox isolation as a swappable interface across runtimes (in-process / WASM / remote) and languages (JS + real Python).

Compose, don't switch (this is the headline pitch — 2026-06-17)

If you already use Vercel AI SDK 6, Mastra, OpenAI Agents JS, Claude Agent SDK, or Cloudflare's codemode — you do not need to switch frameworks to use wasmagent's kernels. Each kernel is an import away:

• Use kernels with Vercel AI SDK
• Use kernels with Mastra
• Use kernels with Claude Agent SDK
• Use kernels with OpenAI Agents JS
• Coming Q3: drop kernel-quickjs / kernel-pyodide into Cloudflare codemode's DynamicWorkerExecutor socket — the architecture document is at docs/strategy/upstream-prs/cloudflare-codemode-byo-executor.md.

The single sentence: swap the executor, keep your framework.