OpenAI-compatible clients generally use
https://api.aixy-gateway.com/v1. Clients such as
Claude Code that append the versioned path use the host https://api.aixy-gateway.com.
Follow the relevant integration guide for the exact client setting.
Decision models evaluate predicates, choices and ordered rubrics through
the same project identity. They use separate typed decision formats and require a compatible
decision provider and model.
Native requests and translation
When the client protocol supports the model and the request’s features, Aixy preserves the native request and response behavior, including streaming events. Otherwise, Aixy translates supported stateless conversations and function-tool exchanges to the model’s execution API and returns the requested client format. Aixy also checks model-specific tool requirements. When an OpenAI model needs Responses for the requested function tools and reasoning, Aixy uses Responses with the same reasoning effort and returns Chat Completions to a Chat Completions client. You can keep the agent’s Aixy connection and change the model without changing its API mode. Activity shows the client API, the provider API selected by Aixy, and the reason for each recorded attempt. Translated Chat Completions uses the conversation and tool history supplied by the client on each call. Reasoning remains enabled at the requested effort, but opaque provider reasoning state is not retained between these calls. Use native Responses when an application needs persisted reasoning or other Responses-specific conversation state. Clients may include plaintextreasoning or reasoning_content in previous assistant messages.
Translation accepts these auxiliary history fields and keeps the visible replies, tool calls and
results. When translating to Responses, Anthropic Messages or Bedrock Converse, Aixy omits the
plaintext history fields: they cannot recreate the destination’s signed or encrypted reasoning
state. This does not change the reasoning effort requested for the new response. Use a matching
native protocol for provider-specific reasoning continuation.
Choose the native format when using provider-specific capabilities. Translation covers the
supported conversation fields; stored conversation references, hosted tools, and advanced fields
require a matching execution contract. Unsupported combinations return an error instead of
silently dropping requested behavior.
For Bedrock Runtime models and inference profiles with verified native Messages support, Aixy
preserves the Messages protocol, including provider-supported tools, caching, thinking and beta
features. The API key selects the project, while the destination model determines feature support;
the same rules apply to SDKs, coding agents and other applications. Reviewed Claude models without
the direct Messages API, and verified application profiles, preserve Messages through Bedrock’s
InvokeModel envelope. Aixy keeps content blocks, tool exchanges and thinking signatures in their
original format and streams events incrementally. Other conversation protocols use their matching
API or supported translation.
Token counting uses the destination’s native counting API. Aixy adapts structured Messages input
for Anthropic, OpenAI, Gemini, Vertex Gemini/Claude and Bedrock, including instructions, tools and supported
multimodal blocks. Bedrock profiles retain their reviewed foundation-model identity for counting.
Provider model, region and permission restrictions still apply. Unsupported input or an unavailable
counter returns an error; Aixy does not substitute a different local tokenizer or generate a response
to estimate its size. Counts describe input size and are separate from billable generation usage.
Anthropic-compatible counting endpoints, including a configured Bedrock Mantle destination, retain
their native structured input. Mantle signing remains restricted to the configured region and the
counting operation; Aixy does not silently move requests to another service or region.
Bedrock Runtime imposes stricter request-metadata limits. For native Messages, Aixy keeps compatible
metadata.user_id values unchanged and represents valid opaque identifiers outside AWS’s character
or length limits as stable SHA-256 identifiers. This affects the identifier in provider-side logs;
conversation, tool and thinking fields retain their native format. The same rule applies to every
client and does not change gateway authentication or authorization.
Beta headers also depend on the destination. On reviewed Bedrock Claude Messages routes, Aixy removes unused
announcements for advisor-tool-2026-03-01, prompt-caching-scope-2026-01-05, and
redact-thinking-2026-02-12. Other beta values are preserved. If the request actually uses an
advisor tool, cache scope, or thinking redaction that cannot be preserved, Aixy returns an explicit compatibility
error. Select a destination supporting that feature or disable it in the client. These rules
apply to every calling application and preserve the Messages response and streaming format.
An explicit omitted thinking display, or a reviewed omitted default, remains unchanged when an
unused redaction announcement is removed. Unknown beta values are forwarded for provider validation.
The API Reference tab is the source for endpoint fields and errors. Model discovery tells you
which IDs your project can address; it does not promise every provider-specific option for every
model.
Configure and verify a client
- Set the gateway base URL and Aixy API key.
- Select an exact model ID or routing alias.
- Supply the conversation in the client’s format and a suitable output-token bound.
- Test text, streaming, and any function tools your application uses.
- Inspect Activity for the resolved provider, attempts, errors, and usage evidence.