Benchmarks

Measured. Reproducible. Caveated.

July 4, 2026 · Node 22 · fresh installs per stack, each with the packages its own docs require for a durable, approvable agent. Script in the repo: bench/footprint.sh.

Dependency packages

Every package is supply-chain surface, and 2025-26 saw real attacks ride trojanized AI-tooling packages.

⬢ nexus0

langgraph63

openai agents106

temporal158

mastra232

inngest agentkit314

Installed footprint

⬢ nexus~1 MB

openai agents64 MB

langgraph90 MB

mastra143 MB

inngest agentkit159 MB

temporal270 MB

Cold-start import

⬢ nexus74 ms

temporal200 ms

openai agents269 ms

langgraph291 ms

inngest agentkit314 ms

mastra505 ms

Moving parts at runtime

Nexus. 1

Your Node process + a file or SQLite. That's the deployment diagram.

OpenAI Agents SDK. 2–4, DIY

Your process + a datastore you build + resume plumbing you build. Their docs recommend adding Temporal/Dapr/Restate for real durability.

LangGraph. 2–4

Your process + Postgres checkpointer; the platform server adds Redis + a license key. Approval transport is yours to build.

Temporal. 3 dev / 5+ prod

Superb durability, via a server, database, UI, admin tools, Elasticsearch, and workers, or their cloud. No budgets either way.

What we don't claim: we don't make the model produce tokens faster. "Faster" = install, import, audit, operate, with fewer components and native budgets, which our July 2026 research could not find in any other open runtime we tested.

How enterprises use this →