.
Don't let avoidable LLM API errors reduce your revenue
Help your users reducing their failed LLM calls with Manifest Autofix.
- OEM / Embed
- White-label ready
- Pay-per-fix
Your team should not debug other companies' APIs
Give users their own keys and providers, and more of their requests break. When one does, they blame you, not the model.
That is exactly where Manifest Autofix shines: A deprecated model name resolves to its successor. An out-of-range parameter is corrected. An oversized payload is trimmed. A malformed tool call is repaired against the schema the developer already declared. Everything is patched on the fly, so agents, workflows and AI applications don't suffer any downtime.
Trusted by engineers who ship at
Your users' failed LLM calls are reducing your product value. We make them disappear
Deprecated / unsupported model
The model you're calling has been retired by the provider. Requests fail until you switch to a supported one.
Malformed parameter
You're sending a parameter the model doesn't accept, like max_tokens instead of max_completion_tokens. Each provider has its own naming.
Malformed message structure
The conversation format doesn't match what the provider expects. Missing fields, wrong roles, or thinking content not included in the response.
Malformed tool schemas
Your tool definitions use schema fields the provider doesn't support, like additionalProperties on Google's API.
Context length exceeded
Your prompt plus the requested output is longer than the model's context window. Something needs to be trimmed.
Rate limit / Insufficient balance
Your API key has hit its usage cap or run out of credits. The provider rejects every request until the limit resets or you top up.
Two ways to run Manifest
Manifest as a sidecar
Your traffic never leaves your stack. When an LLM API call fails, you send the failure to Manifest and we send back the fix to apply.
- Out of the request path, successful calls go straight to the provider
- Fixes directly the request or suggests the change to do
- Full bodies not sent to Manifest
Manifest as a proxy
Every LLM call runs through Manifest AI model gateway, combining fallback and Autofix to ensure the maximum success rate.
- One API endpoint for all models
- Custom routing and fallback
- Observability and full traces
Embed Manifest inside your product.
Frequently asked questions
How do we embed Manifest?
Two ways. As a sidecar: when a provider rejects one of your users' calls, you send us the error and we send back the patch to apply, so successful traffic never touches us. As a proxy: calls run through the Manifest gateway and come back fixed, with routing, fallback and full traces.
Do our users' payloads reach your servers?
In sidecar mode, no. You send the error and the fields it points at, we return the correction. In proxy mode the request transits the gateway like any other Manifest call. Pick the mode your security team can sign off on.
Which errors does Autofix actually cover?
Today it covers a first set of known errors, each with a known correction: renamed parameters, retired model names, values a model refuses. That set grows as new errors show up in traffic.
Every attempt is logged with the error the provider returned and the exact patch applied, so your support team sees the cause instead of guessing. Frequency data per provider lives in our public error catalog.
Can we white-label it?
Yes. Ship it as "powered by Manifest", or with nothing on it but your own name. We settle the branding question on the partner call.
What does the design-partner program include?
Early access to the embed API before it opens up, a direct line to the people building it, and real weight on what ships next. We keep the group small so the feedback loop stays short.
How is pricing handled for our users?
You pay per fix, not per request and not per seat. Rates are still being set with design partners, so we work them out case by case instead of publishing a number we would have to walk back.