Providers

Every request is bring-your-own-key: Zerocache never holds a provider credential beyond the duration of one request. {provider} in the URL path selects which adapter handles the request; the three cloud embedding providers encode routing coordinates (region, project, input type) into the model string itself rather than a new wire field, since model is already free-form per-request and already lands in the cache key.

Chat completions

POST /{provider}/v1/chat/completions — nine hosted OpenAI-wire endpoints, registered with zero config. Add or repoint one (self-hosted vLLM / Ollama, an internal gateway) with ZEROCACHE_CHAT_PROVIDERS="name=url,…", where the URL is the prefix up to but not including /chat/completions.

openaimistralgeminigroqdeepseektogetheropenrouterxaifireworks

Anthropic Messages

POST /{provider}/v1/messages — Claude's native wire shape. Built-inanthropic → https://api.anthropic.com; add or repoint withZEROCACHE_MESSAGES_PROVIDERS="name=url,…" (bare-origin values — the adapter appends /v1/messages). Authorization: Bearer is rewritten tox-api-key upstream.

Embeddings

OpenAI

auth: Bearer key
text: yes
images: no

Mistral

auth: Bearer key
text: yes
images: no

Gemini

auth: Bearer key
text: yes
images: yes

HuggingFace

auth: Bearer key
text: yes
images: no

model is part of the URL path, not the body

Bedrock

auth: Bearer key (AWS bearer API key)
text: yes
images: no

[<region>/]<modelId>[#<input_type>]

Vertex AI

auth: Bearer (GCP OAuth2 token)
text: yes
images: no

[<location>/<project>/]<modelId>[#<task_type>]

Azure

auth: Bearer (Entra token, default) or api-key
text: yes
images: no

[foundry:]<deployment>[#<input_type>]

Gemini is the only provider with image embeddings, and never reports token usage — usage is always zero on its responses.
Bedrock authenticates with an Amazon bearer API key — no AWS SigV4. Titan and Cohere sit behind one router.
Vertex AI uses a GCP OAuth2 access token, which lives about an hour; since owner_id is a hash of that credential, a Vertex caller's cache namespace rotates when the token does — fine for one long ingestion run, not across days.
Azure registers as a provider only once ZEROCACHE_AZURE_OPENAI_BASE_URL or ZEROCACHE_AZURE_FOUNDRY_BASE_URL is set; with neither, /azure/v1/embeddings 404s. It covers the Azure OpenAI GA path and Foundry Models.
OpenAI and Mistral take a configurable base URL, so a self-hosted wire-compatible endpoint (vLLM, LM Studio, TGI) works too. The cloud adapters are mock-tested only — no live-key smoke test in this environment.