Full voltage.
Full control.
Rheostat is the runtime control layer for AI releases. Roll out, dial back and self-heal every change in production — without a redeploy.
14-day trial · No card · SOC 2 Type II
ROLLOUT · checkout-v2
- 5%
- 25%
- 50%
- 100%
HALLUCINATION RATE · 7d
GUARDRAIL
Armed · 3 flags watched
Every change,
on a dial.
Three controls, one runtime. Decide who sees what, how fast, and what happens when it goes wrong — without shipping new code.
Roll out by the notch.
Release to 5%, then 25%, then everyone. Rheostat watches the numbers between each notch and holds the dial if they slip.
Learn more →Rollout schedule
Guardrails for agents in production.
Judge every agent response against the checks you choose. Steer, swap models or pause the agent the moment a check fails.
Learn more →Add judge
- AccuracyFactual correctness against sourcesOn
- ToxicityHarmful or abusive languageOff
- PII leakageEmails, cards, national IDsOn
Self-heal before anyone gets paged.
Write the rule once. When latency, errors or cost cross the line, Rheostat dials the change back to zero on its own.
Learn more →Incident · checkout-v2
Recovered- 14:02:07p95 latency 912 ms
- 14:02:09Rule fired → dial to 0%
- 14:03:41p95 back to 214 ms
AI that behaves.
However you build.
- 3.1B
- flag evaluations a day
- <20 ms
- p99 rule evaluation at the edge
- 99.99%
- uptime, measured over 12 months
Keep agents
inside the lines.
Evaluate every response, trigger on what matters and roll out new prompts and models by percentage — the same controls you already trust for code.
Online evals
Score live traffic with judges you define. Sample 1% or 100%.
Offline evals
Replay a dataset against a new prompt before a single user sees it.
Adaptive triggers
Fire a rollback, a model swap or a page when a metric crosses a line.
Progressive rollouts
Move prompts and models from 5% to 100% on a schedule you control.
Built for developers
and their agents.
Learn the patterns.
Guides for canary releases, kill switches and eval-driven prompts.
Go agent-native.
Let your coding agent create, flip and clean up flags through our MCP server.
Copy, paste, go.
Drop this prompt into your coding assistant and have your first flag running in under a minute.
Set up Rheostat: npx rheostat init --agent && rheostat flag create new-checkout --rollout 5%“We stopped scheduling releases around the on-call rota. Now we ship on Friday afternoon and let the dial do the worrying.”
- 4×
- more releases a week
- 11 min
- median time to recover
- 0
- release-night pages in Q3
“The release freeze is gone. Every service ships behind a dial and we roll back in seconds, not war rooms.”
- 900
- services on percentage rollouts
- 0
- freeze windows this year
- 6 s
- median rollback
“We swap models per customer tier without a deploy. Judges tell us before users do when one drifts.”
- 3
- models behind one flag
- −22%
- inference cost
- 99.2%
- judge agreement
“Prompt changes used to wait for the next release train. Now they go out at 5% the same hour they are written.”
- 1 h
- prompt change to 5%
- 40
- experiments a month
- 0
- hot fixes in Q3
Resources to
get you started.
Turn it up.
Start with one flag on one service. Most teams dial their first release in the same afternoon.




