Skip to main content

How It Works

Your server sends a provider-native payload to a regional Makas mirror. Makas authenticates the regional service credential, validates the provider credential, applies eligible cache/routing behavior, forwards using native provider auth and records usage metadata.

ProviderMirror base
OpenAIhttps://eu.makas.ai/proxy/openai/v1
Anthropic SDKhttps://eu.makas.ai/proxy/anthropic (SDK adds /v1/messages)
Anthropic HTTPhttps://eu.makas.ai/proxy/anthropic/v1
Googlehttps://eu.makas.ai/proxy/google/v1beta

The regional selector changes runnable code and action links. Base URLs in static prose are EU examples; use the complete region-aware quickstarts when copying a request.

Routing stays within the provider selected by the proxy path. It evaluates applicable customer rules before global defaults, with lower priority numbers evaluated first. The global low-confidence guard can preserve the requested model. Capabilities, request shape and STICKY conversation continuity can prevent a change. A successful request can have Makas-Was-Routed: false and zero savings.

GroupDefault behavior
COST_DOWN_CHEAPESTOn: simple mechanical tasks to cheapest capable models
COST_DOWN_BUDGETOn: light routine tasks to budget models
COST_DOWN_MIDOn: economical models for eligible everyday and selected complex tasks
CODE_SPECIALIZATIONOn: eligible code tasks to purpose-built models at the same price
MODERNIZEOn when the plan includes automatic model upgrades: same or lower cost
RETIRE_DEPRECATEDOff: may raise cost; current opt-out API cannot enable it
QUALITY_UPGRADEOff: raises cost; current opt-out API cannot enable it

Default-on groups can be opted out in Routing Rules. Removing a cost-raising group from an opt-out list does not turn it on; there is no current customer enable API for those groups. MODERNIZE requires Growth+ in the planned paid tier system and is included in Founding.

Custom rules require Growth+ when billing is enabled; Founding includes up to 10. Per-key routing opt-outs require Scale+ in the planned paid tiers and are included in Founding. Do not assume all rules apply to every request.

Compare model requested/used and the optional rule in Request Logs. Conditional response headers can also help for buffered inference.

Semantic caching is disabled for the first launch. Founding does not include semantic caching. The exact-request cache is a separate in-memory optimization for eligible identical requests; its eligibility depends on request shape and provider cache directives. Do not rely on a cache hit for correctness or assume every request is cached.

There is no customer Settings → Caching control. These docs do not promise a cache-hit response header.

If semantic caching is explicitly enabled in a future deployment, eligible Growth+ requests send prompts to OpenAI for embeddings, retain hashes and unencrypted embeddings, and store encrypted responses per customer with enforced retention. There is no separate customer opt-out. This is distinct from ordinary routing, which does not persist prompt plaintext.

See privacy and security and response headers.