Now in public beta · durable agent runtime

Ship agents that don't break in production.

Nebula runs, traces, and guards your AI agents in production. Generation is the easy half, Nebula handles the hard half: verification.

nebula · live trace
  • run.startcheckout-agent · run_8f21
  • tool.calllookup_order(1187)
  • modelclaude · 412 tokens · 41ms
  • guardrailrefund > $200 · needs approval
  • humanapproved by @mara
  • run.okresolved in 1.2s

Trusted by teams at

NorthwindSableCobaltMeridianHalo
  • Northwind
  • Sable
  • Cobalt
  • Meridian
  • Halo
  • Driftwood
  • Verge
  • Loop

The platform

Everything around the agent, so the agent is the easy part.

Durable runs

Every agent run is checkpointed and resumable. A crash, a redeploy, or a flaky tool never loses state.

Live tracing

See every step, tool call, and token as it happens. Replay any run against a new prompt or model.

Guardrails

Policy checks between generation and action. Risky calls pause for a human instead of firing blind.

Evals in CI

Score every change against a suite. The build fails when your agent gets measurably dumber.

Scoped tools

Expose exactly the tools each agent needs, with per-client permission boundaries and an audit log.

99.98%
runtime uptime
41ms
p50 step latency
2.1B
agent runs served
8,400
teams building

How it works

From local script to production, in three moves.

01

Define

Wrap your agent in a few lines. Declare its tools, its guardrails, and what a good run looks like.

02

Deploy

Push to Nebula. We run it durably at the edge, checkpoint every step, and scale it to zero when idle.

03

Observe

Watch live traces, catch regressions in CI, and let humans approve the calls that matter.

Teams sleep better with Nebula.

“We deleted an entire internal platform team. Nebula runs our agents, traces them, and tells us before something breaks.”
Priya Nair
VP Engineering, Sable
“The guardrails alone paid for it. Our refund agent pauses on anything above policy and a human clears it in one tap.”
Marcus Feld
Founder, Driftwood
“Evals in CI changed how we ship prompts. No more "it felt better", the build tells us if it got worse.”
Ana Costa
Staff ML Engineer, Meridian

Pricing

Start free. Scale when your agents do.

Hobby

For side projects and first agents.

$0forever
  • 10k runs / month
  • Live tracing
  • Community support
  • 1 project
Start free
Most popular

Team

For teams shipping agents to production.

$49per seat / month
  • Unlimited runs
  • Guardrails + human-in-the-loop
  • Evals in CI
  • Audit logs & SSO
  • Priority support
Start 14-day trial

Scale

For platforms running agents at scale.

Customannual
  • Dedicated regions
  • Custom SLAs
  • On-prem option
  • Solutions engineer
Talk to us

Questions, answered.

What exactly does Nebula run?+

Any agent you can express as code, LangGraph, the Vercel AI SDK, a raw tool-calling loop, or your own framework. Nebula handles durability, tracing, guardrails, and evals around it.

Do I have to change my model provider?+

No. Nebula is provider-agnostic. Point it at Claude, GPT, Gemini, or a local model, and swap freely without touching your run logic.

How does human-in-the-loop work?+

You mark which actions need approval. Nebula pauses the run, notifies a reviewer on web or mobile, and resumes cleanly with their decision injected.

Is my data used for training?+

Never. Your prompts, traces, and tool data stay yours. We are SOC 2 Type II and offer an on-prem option on Scale.

Can I self-host?+

Yes, on the Scale plan. The same runtime, tracing, and guardrails run inside your own cloud.

Put your agents somewhere they can't quietly fail.