Your AI stays online,
even during outage

Behind the scenes

OpenAI · gpt-4ostopped answering

Anthropic · claude-sonnet-4-5answered · +59 ms

What your user sees

I ordered the black jacket last week.That's order 4821. It ships tomorrow morning.Can you send it to my office instead?Done. It'll arrive at your office on Thursday.

OpenAI
Anthropic
Gemini
MistralSoon
MiniMax
OpenAI
Anthropic
Gemini
MistralSoon
MiniMax
OpenAI
Anthropic
Gemini
MistralSoon
MiniMax
OpenAI
Anthropic
Gemini
MistralSoon
MiniMax

Why Metriqual

Every model call you make runs through providers you do not control. Metriqual holds the keys, picks the model, replays the request when one stops answering, and prices every call against the key that made it.

See which model costs you less.

Reliability

What we do with every call you send

A failing provider is taken out, not retried into the ground

openaiopenerror rate over the window · out for 30s
geminihalf-opencooldown done · one probe through
anthropicclosedserving

A Rust proxy in the path

Axum on Tokio. Provider bytes are forwarded as they arrive, so streaming stays streaming.

rust
axum
tokio

Spend per key

openai/gpt-4o$18.40

anthropic/claude-sonnet-4-5$11.02

openai/gpt-4o-mini$2.90

Every attempt on the record

1openai/gpt-4otimeout
2gemini/gemini-2.5-proskipped_circuit_open
3anthropic/claude-sonnet-4-5success

Detectors that read the answer, not the status code

history_integrityThe stored conversation and the one that came back stopped agreeing.
toolcall_arguments_corruptThe model returned tool arguments that are not valid JSON.
recorded, not swallowedBoth land on the request record and on a response header you can read.

Failovers, recoveries and request logs go where your team already watches.

Slack
Discord
Microsoft Teams
Datadog
Splunk
OpenTelemetry
Your own endpoint
Slack
Discord
Microsoft Teams
Datadog
Splunk
OpenTelemetry
Your own endpoint
Slack
Discord
Microsoft Teams
Datadog
Splunk
OpenTelemetry
Your own endpoint
Slack
Discord
Microsoft Teams
Datadog
Splunk
OpenTelemetry
Your own endpoint

Modalities

Chat, voice, image, video. Same key.

Run all four in the playground against the live agent before anything ships: compare models, watch per-stage latency, keep the result. Failover covers every one of them.

Search agents…

support-bot

OpenAI

Anthropic

Knowledge base · 3 docs

Ask support-bot something…

Live web search

10 decisions sit between your app and production.

Provider chain
Failover order
Stored, passthrough or managed keys
PII redaction
Content filters
Model allowlist
System prompts
Prompt versioning
Conversation state
Webhooks and log drains

We build all 10 for you.

You paste this. Two lines, whichever you write in.

curl

https://api.metriqual.com/v1/chat/completions

-H "Authorization: Bearer mql_your_key"

typescript

baseURL: "https://api.metriqual.com/v1"

apiKey: process.env.METRIQUAL_API_KEY

python

base_url="https://api.metriqual.com/v1"

api_key=os.environ["METRIQUAL_API_KEY"]

Request path

How it works

One request, end to end: your app, the routing decision, the provider that answers, and the user waiting on it.

Book a setup call Read the docs

Your app sends the request
The same SDK call you write today, pointed at one base URL and one key. Nothing else in the call changes.
Metriqual routes it
We hold the provider keys and pick the model by the rules you set, so your code never carries a provider name.
The provider answers, or the next one does
If a provider stops answering, the request is replayed against the next in the chain, carrying its conversation history.
Your user gets the response
It comes back priced: model, key and spend attached to the request, and a key cannot spend past the ceiling you set.