providers.yaml

# SPDX-License-Identifier: Apache-2.0
# Copyright (C) 2026 Busbar Inc and contributors
#
# Verified provider catalog for busbar.
#
# Each entry maps a provider NAME to its wire `protocol`, `base_url`, and an optional `error_map`
# (provider-specific JSON error code -> canonical disposition: billing | rate_limit). Operators
# deploy by referencing a provider name in config.yaml and supplying its key via an env var.
#
# busbar's thesis is "protocols, not providers": it implements a small set of wire protocols
# (anthropic, openai, gemini, bedrock, responses, cohere) losslessly, and any provider speaking one of
# those protocols is just a catalog entry here - no new code. A client speaking any protocol can
# target any provider; busbar translates through its superset IR when they differ.
#
# Auth is a per-protocol seam, not a single scheme: most providers use `Authorization: Bearer <key>`,
# gemini uses the `x-goog-api-key` header, and bedrock signs each request with AWS SigV4 (see below).
# Azure OpenAI is supported today via the `auth: api-key` override (a fill-in template is below); Google
# Vertex AI via the `auth: jwt-bearer` OAuth adapter (a fill-in template is below the gemini entry).
# HTTP-status errors (429/5xx/401/...) are classified by the circuit breaker without an error_map;
# error_map is only for provider-specific JSON codes.

# -- Anthropic protocol (/v1/messages) -----------------------------------------------------------
anthropic:
  protocol: anthropic
  base_url: https://api.anthropic.com
  error_map: {}

# z.ai exposes two surfaces. `z.ai` is the Anthropic-compatible coding-agent endpoint
# (/api/anthropic + /v1/messages). `zai-api` (below, in the OpenAI section) is the general
# pay-as-you-go API (OpenAI-compatible, version in the base path). Pick the one your key targets.
z.ai:
  protocol: anthropic
  base_url: https://api.z.ai/api/anthropic
  error_map:
    "1113": billing
    "1302": rate_limit

# -- OpenAI Chat Completions protocol (/v1/chat/completions) --------------------------------------
# Any provider exposing an OpenAI-compatible chat-completions endpoint. The `base_url` is composed
# with `/v1/chat/completions`; the entries below place each provider's prefix accordingly.
openai:
  protocol: openai
  base_url: https://api.openai.com
  error_map:
    insufficient_quota: billing
    rate_limit_exceeded: rate_limit

# Oracle OCI Generative AI, the OpenAI-compatible surface (/openai/v1/chat/completions), serving
# OCI's hosted catalog (OpenAI, Meta Llama, xAI, Google, and Cohere models). Since Jan 2026 OCI
# issues plain API keys (Bearer), so NO OCI request-signing is needed - set api_key_env to an env var
# holding the key. REGIONAL: the region is in the host - set it to yours (`us-chicago-1` /
# `us-ashburn-1` carry the broadest model set). Throttling is HTTP 429 (classified by the breaker);
# OCI can also return a 400 carrying `TooManyRequests`/`QuotaExceeded`/`LimitExceeded`, and the
# OpenAI-compat layer may emit OpenAI-style codes - both are mapped. (An OCI Responses transport also
# exists at /openai/v1/responses; add a `protocol: responses` entry if you want it.)
oci-genai:
  protocol: openai
  base_url: https://inference.generativeai.us-chicago-1.oci.oraclecloud.com/openai
  error_map:
    TooManyRequests: rate_limit
    QuotaExceeded: billing
    LimitExceeded: billing
    rate_limit_exceeded: rate_limit
    insufficient_quota: billing

groq:
  protocol: openai
  base_url: https://api.groq.com/openai
  error_map:
    rate_limit_exceeded: rate_limit

# z.ai general API (pay-as-you-go), OpenAI-compatible. The version lives in the base path
# (/api/paas/v4) and the endpoint is <base>/chat/completions, so the appended path is overridden.
zai-api:
  protocol: openai
  base_url: https://api.z.ai/api/paas/v4
  path: /chat/completions
  error_map:
    "1113": billing
    "1302": rate_limit

together:
  protocol: openai
  base_url: https://api.together.xyz
  error_map:
    rate_limit_exceeded: rate_limit

# Fireworks signals errors by HTTP status only (docs.fireworks.ai/guides/inference-error-codes);
# live error bodies carry gRPC-style codes (NOT_FOUND), so no JSON code map applies.
fireworks:
  protocol: openai
  base_url: https://api.fireworks.ai/inference
  error_map: {}

# DeepSeek signals errors by HTTP status only (api-docs.deepseek.com/quick_start/error_codes):
# 402 = insufficient balance, 429 = rate limit. No machine-readable code strings documented.
deepseek:
  protocol: openai
  base_url: https://api.deepseek.com
  error_map: {}

mistral:
  protocol: openai
  base_url: https://api.mistral.ai
  error_map:
    rate_limit_exceeded: rate_limit

xai:
  protocol: openai
  base_url: https://api.x.ai
  error_map:
    rate_limit_exceeded: rate_limit

# Codes per openrouter.ai/docs/api/reference/errors-and-debugging: 402 payment_required
# (insufficient credits), 429 rate_limit_exceeded.
openrouter:
  protocol: openai
  base_url: https://openrouter.ai/api
  error_map:
    payment_required: billing
    rate_limit_exceeded: rate_limit

cerebras:
  protocol: openai
  base_url: https://api.cerebras.ai
  error_map:
    rate_limit_exceeded: rate_limit

deepinfra:
  protocol: openai
  # base_url embeds the API version (/v1/openai) and serves /chat/completions directly, so override
  # the appended path to avoid a doubled /v1.
  base_url: https://api.deepinfra.com/v1/openai
  path: /chat/completions
  # DeepInfra signals rate limits as plain HTTP 429 (docs.deepinfra.com/account/rate-limits);
  # no JSON code strings documented.
  error_map: {}

# Codes per platform.kimi.ai/docs/api/errors.md (Moonshot's own error types, not OpenAI's).
moonshot:
  protocol: openai
  base_url: https://api.moonshot.ai
  error_map:
    rate_limit_reached_error: rate_limit
    engine_overloaded_error: rate_limit
    exceeded_current_quota_error: billing

# Additional OpenAI-compatible providers (base_urls verified against provider docs). Most signal
# errors via HTTP status (classified by the breaker), so error_map is empty; add billing/rate
# codes per provider only when documented. `path` overrides are set where the version lives in the
# base path and the endpoint is <base>/chat/completions.
meta-llama:
  protocol: openai
  base_url: https://api.llama.com/compat
  error_map: {}

nvidia-nim:
  protocol: openai
  base_url: https://integrate.api.nvidia.com
  error_map: {}

sambanova:
  protocol: openai
  base_url: https://api.sambanova.ai
  error_map: {}

hyperbolic:
  protocol: openai
  base_url: https://api.hyperbolic.xyz
  error_map: {}

# novita.ai/docs/api-reference: base https://api.novita.ai/openai, standard /v1 path.
novita:
  protocol: openai
  base_url: https://api.novita.ai/openai
  error_map: {}

# Nebius AI Studio became Nebius Token Factory (Nov 2025); docs.tokenfactory.nebius.com
# documents only the new host.
nebius:
  protocol: openai
  base_url: https://api.tokenfactory.nebius.com
  error_map: {}

featherless:
  protocol: openai
  base_url: https://api.featherless.ai
  error_map: {}

friendliai:
  protocol: openai
  base_url: https://api.friendli.ai/serverless
  error_map: {}

scaleway:
  protocol: openai
  base_url: https://api.scaleway.ai
  error_map: {}

ovhcloud:
  protocol: openai
  base_url: https://oai.endpoints.kepler.ai.cloud.ovh.net
  error_map: {}

nscale:
  protocol: openai
  base_url: https://inference.api.nscale.com
  error_map: {}

gmi:
  protocol: openai
  base_url: https://api.gmi-serving.com
  error_map: {}

baseten:
  protocol: openai
  base_url: https://inference.baseten.co
  error_map: {}

chutes:
  protocol: openai
  base_url: https://llm.chutes.ai
  error_map: {}

# Alibaba DashScope (Qwen) international host; the version lives in /compatible-mode/v1 so the
# default /v1/chat/completions composes correctly.
dashscope:
  protocol: openai
  base_url: https://dashscope-intl.aliyuncs.com/compatible-mode
  error_map: {}

minimax:
  protocol: openai
  base_url: https://api.minimax.io
  error_map: {}

sarvam:
  protocol: openai
  base_url: https://api.sarvam.ai
  error_map: {}

# Multi-model aggregators exposing one OpenAI-compatible chat-completions endpoint.
aimlapi:
  protocol: openai
  base_url: https://api.aimlapi.com
  error_map: {}

cometapi:
  protocol: openai
  base_url: https://api.cometapi.com
  error_map: {}

nano-gpt:
  protocol: openai
  base_url: https://nano-gpt.com/api
  error_map: {}

publicai:
  protocol: openai
  base_url: https://api.publicai.co
  error_map: {}

synthetic:
  protocol: openai
  base_url: https://api.synthetic.new/openai
  error_map: {}

wandb:
  protocol: openai
  base_url: https://api.inference.wandb.ai
  error_map: {}

# Azure OpenAI is the OpenAI chat-completions protocol behind a different auth shape: the `api-key`
# header (not bearer), with the API version and deployment in the path. Its host and deployment are
# per-resource, so it is not shipped as a live entry - copy this template into your config.yaml
# providers map and fill in your values:
#
#   azure:
#     protocol: openai
#     base_url: https://YOUR-RESOURCE.openai.azure.com
#     auth: api-key                                  # send `api-key: <key>` instead of bearer
#     path: /openai/deployments/YOUR-DEPLOYMENT/chat/completions?api-version=2024-06-01
#     error_map: {}
#     api_key_env: AZURE_OPENAI_KEY
#
# Azure OpenAI via Microsoft Entra ID (AAD) instead of an API key: same OpenAI protocol + deployment
# path, but authed with `auth: oauth-client-credentials` — busbar exchanges an app registration's
# client_id:client_secret at the tenant token endpoint for an auto-refreshed bearer. Tenant + app are
# per-account, so it is a fill-in template. `api_key_env` holds `client_id:client_secret`:
#
#   azure-entra:
#     protocol: openai
#     base_url: https://YOUR-RESOURCE.openai.azure.com
#     path: /openai/deployments/YOUR-DEPLOYMENT/chat/completions?api-version=2024-06-01
#     auth: oauth-client-credentials
#     token_url: https://login.microsoftonline.com/YOUR-TENANT/oauth2/v2.0/token
#     scope: https://cognitiveservices.azure.com/.default
#     error_map: {}
#     api_key_env: AZURE_ENTRA_CREDS                 # value = "client_id:client_secret"


# --- Added 2026-07-09: docs-verified against each provider's own API documentation ---------------

# DigitalOcean Gradient serverless inference.
digitalocean:
  protocol: openai
  base_url: https://inference.do-ai.run
  error_map:
    too_many_requests: rate_limit

# GitHub Models inference endpoint (no /v1 segment in the documented path).
github-models:
  protocol: openai
  base_url: https://models.github.ai/inference
  path: /chat/completions
  error_map: {}

# Hugging Face Inference Providers router.
huggingface:
  protocol: openai
  base_url: https://router.huggingface.co
  error_map: {}

# SiliconFlow international endpoint. The mainland-China host is api.siliconflow.cn.
siliconflow:
  protocol: openai
  base_url: https://api.siliconflow.com
  error_map: {}

# Upstage Solar.
upstage:
  protocol: openai
  base_url: https://api.upstage.ai
  error_map: {}

# StepFun international endpoint. The mainland-China host is api.stepfun.com.
stepfun:
  protocol: openai
  base_url: https://api.stepfun.ai
  error_map: {}

# Zhipu bigmodel.cn (mainland-China GLM API; the international surface is z.ai, above).
# Numeric JSON codes per their error reference: 1113 arrears / 1309 plan expired,
# 1302/1305/1308/1310/1313 rate and usage limits.
zhipuai:
  protocol: openai
  base_url: https://open.bigmodel.cn/api/paas/v4
  path: /chat/completions
  error_map:
    "1113": billing
    "1309": billing
    "1302": rate_limit
    "1305": rate_limit
    "1308": rate_limit
    "1310": rate_limit
    "1313": rate_limit

# Xiaomi MiMo. Their documented key env convention is MIMO_API_KEY.
xiaomi:
  protocol: openai
  base_url: https://api.xiaomimimo.com
  error_map: {}

# Vultr serverless inference.
vultr:
  protocol: openai
  base_url: https://api.vultrinference.com
  error_map: {}

# Ollama Cloud.
ollama-cloud:
  protocol: openai
  base_url: https://ollama.com
  error_map: {}

# Inception Labs (Mercury diffusion LLMs).
inception:
  protocol: openai
  base_url: https://api.inceptionlabs.ai
  error_map: {}

# Morph (code-apply models).
morph:
  protocol: openai
  base_url: https://api.morphllm.com
  error_map: {}

# Meituan LongCat. The documented endpoint already ends in /v1/chat/completions.
longcat:
  protocol: openai
  base_url: https://api.longcat.chat/openai
  error_map:
    insufficient_quota: billing
    rate_limit_exceeded: rate_limit

# Sakana AI (fugu models).
sakana:
  protocol: openai
  base_url: https://api.sakana.ai
  error_map: {}

# NEAR AI cloud.
nearai:
  protocol: openai
  base_url: https://cloud-api.near.ai
  error_map: {}


# --- Added 2026-07-09: verified against each provider's own API documentation ---

# Evidence: doc-en.302.ai documents POST https://api.302.ai/v1/chat/completions, Bearer auth.
302ai:
  protocol: openai
  base_url: https://api.302.ai
  error_map: {}

# Evidence: abacus.ai/help/developer-platform/route-llm/ gives base https://routellm.abacus.ai/v1, Bearer.
abacus:
  protocol: openai
  base_url: https://routellm.abacus.ai
  error_map: {}

# Evidence: abliteration.ai/docs/openai-compatibility documents /v1/chat/completions, Bearer (ak_ keys).
abliteration-ai:
  protocol: openai
  base_url: https://api.abliteration.ai
  error_map: {}

# Evidence: desktop.ambient.xyz/reference/api-endpoints/ shows curl to /v1/chat/completions with Bearer $AMBIENT_KEY.
ambient:
  protocol: openai
  base_url: https://api.ambient.xyz
  error_map: {}

# Evidence: docs.anyapi.ai/guides/errors.md documents 402 budget_exceeded, 429 rate_limit_error.
anyapi:
  protocol: openai
  base_url: https://api.anyapi.ai
  error_map:
    budget_exceeded: billing
    rate_limit_error: rate_limit

# Evidence: docs.auriko.ai/contract/error-codes enumerates all three codes.
auriko:
  protocol: openai
  base_url: https://api.auriko.ai
  error_map:
    budget_exhausted: billing
    insufficient_quota: billing
    rate_limit_exceeded: rate_limit

# Evidence: alipaytbox.yuque.com/sxs0ba/ling/api_chat documents POST https://api.ant-ling.com/v1/chat/completions, Bearer sk-studio- token.
bailing:
  protocol: openai
  base_url: https://api.ant-ling.com
  error_map: {}

# Evidence: api.berget.ai/openapi.json lists /v1/chat/completions with HTTP bearer securityScheme.
berget:
  protocol: openai
  base_url: https://api.berget.ai
  error_map: {}

# best-effort: codes from native API status reference, not the OpenAI shim
# Evidence: docs.clarifai.com/compute/inference/open-ai/ documents base https://api.clarifai.com/v2/ext/openai/v1 with PAT as Bearer.
clarifai:
  protocol: openai
  base_url: https://api.clarifai.com/v2/ext/openai
  error_map:
    "11003": rate_limit
    "11004": rate_limit
    "11005": rate_limit
    "11006": billing
    "11000": billing

# Evidence: claudin.io/docs/clients/openai-compatible/ shows curl to https://api.claudin.io/v1/chat/completions, Bearer.
claudinio:
  protocol: openai
  base_url: https://api.claudin.io
  error_map: {}

# Evidence: docs.sherlock.cloudferro.com/chat-completion-endpoint/ documents https://api-sherlock.cloudferro.com/openai/v1/chat/completions, Bearer $SHERLOCK_API_KEY.
cloudferro-sherlock:
  protocol: openai
  base_url: https://api-sherlock.cloudferro.com/openai
  error_map: {}

# Evidence: docs.cortecs.ai/quickstart shows exact curl to https://api.cortecs.ai/v1/chat/completions, Bearer.
cortecs:
  protocol: openai
  base_url: https://api.cortecs.ai
  error_map: {}

# Evidence: crof.ai/docs.md documents https://crof.ai/v1/chat/completions, Bearer (note: bare domain, no api. subdomain).
crof:
  protocol: openai
  base_url: https://crof.ai
  error_map: {}

# Evidence: crossmodel.ai/docs/errors documents 402 insufficient_balance/no_wallet, 429 rate_limit_exceeded.
crossmodel:
  protocol: openai
  base_url: https://api.crossmodel.ai
  error_map:
    insufficient_balance: billing
    no_wallet: billing
    rate_limit_exceeded: rate_limit

# Evidence: dinference.com/docs documents base https://api.dinference.com/v1, /v1/chat/completions, DINFERENCE_KEY.
dinference:
  protocol: openai
  base_url: https://api.dinference.com
  error_map: {}

# Evidence: docs.d.run/en/models/api-call documents MaaS base https://chat.d.run/v1, Bearer sk- key. (Dedicated instances use per-region hosts, e.g. sh-02.d.run.)
drun:
  protocol: openai
  base_url: https://chat.d.run
  error_map: {}

# Evidence: docs.evroc.com/products/think/apis/v1/openapi.yaml - server https://models.think.evroc.com/v1, path /chat/completions, bearer. (Public beta; APIs may change.)
evroc:
  protocol: openai
  base_url: https://models.think.evroc.com
  error_map: {}

# Evidence: docs.fastrouter.ai/api-reference/chat-completions - endpoint https://api.fastrouter.ai/api/v1/chat/completions (note extra /api segment).
fastrouter:
  protocol: openai
  base_url: https://api.fastrouter.ai/api
  error_map: {}

# Evidence: docs.frogbot.ai/api-reference - base https://app.frogbot.ai/api/v1, Bearer $FROGBOT_API_KEY.
frogbot:
  protocol: openai
  base_url: https://app.frogbot.ai/api
  error_map: {}

# error_map deliberately omitted: type "request_failed" is not unique to billing ("Insufficient credits"
# arrives as HTTP 429 with type request_failed); message-matching would be needed for precision.
# Evidence: docs.helicone.ai/gateway/integrations/codex.md uses base https://ai-gateway.helicone.ai/v1 with OpenAI clients.
helicone:
  protocol: openai
  base_url: https://ai-gateway.helicone.ai
  error_map: {}

# Evidence: hpc-ai.com/doc .../Create-Chat-Completions/ documents https://api.hpc-ai.com/inference/v1/chat/completions, Bearer.
hpc-ai:
  protocol: openai
  base_url: https://api.hpc-ai.com/inference
  error_map: {}

# Evidence: platform.iflow.cn/en/docs/api-reference (endpoint) + /docs/error-codes (codes; page also covers search API - verify at runtime).
iflowcn:
  protocol: openai
  base_url: https://apis.iflow.cn
  error_map:
    "40303": rate_limit
    "60400": billing

# Evidence: docs.inceptron.io API Reference documents base https://api.inceptron.io/v1, Bearer INCEPTRON_API_KEY.
inceptron:
  protocol: openai
  base_url: https://api.inceptron.io
  error_map: {}

# Evidence: docs.inference.net/api/api-quickstart.md - base https://api.inference.net/v1, Bearer INFERENCE_API_KEY.
inference:
  protocol: openai
  base_url: https://api.inference.net
  error_map: {}

# Evidence: io.net/docs .../get-started-with-io-intelligence-api - endpoint https://api.intelligence.io.solutions/api/v1/chat/completions (note /api prefix and different domain from io.net).
io-net:
  protocol: openai
  base_url: https://api.intelligence.io.solutions/api
  error_map: {}

# Evidence: docs.jiekou.ai/docs/models/reference-error-code.md + FAQ confirming /openai/v1/chat/completions form. (China mirror: api.highwayapi.ai/openai.)
jiekou:
  protocol: openai
  base_url: https://api.jiekou.ai/openai
  error_map:
    NOT_ENOUGH_BALANCE: billing
    RATE_LIMIT_EXCEEDED: rate_limit
    TOKEN_LIMIT_EXCEEDED: rate_limit

# Evidence: kenari.id/docs curl example POST https://kenari.id/v1/chat/completions, Bearer kn- keys, error codes enumerated.
kenari:
  protocol: openai
  base_url: https://kenari.id
  error_map:
    insufficient_balance: billing
    payg_limit_reached: billing
    rate_limit_exceeded: rate_limit
    plan_limit_reached: rate_limit

# Evidence: kilo.ai/docs/gateway/api-reference - endpoint /api/gateway/chat/completions has NO /v1, so path override required. Only entry in this batch needing one.
kilo:
  protocol: openai
  base_url: https://api.kilo.ai/api/gateway
  path: /chat/completions
  error_map: {}

# Evidence: docs.getlilac.com/inference/chat-completions documents POST https://api.getlilac.com/v1/chat/completions, Bearer.
lilac:
  protocol: openai
  base_url: https://api.getlilac.com
  error_map: {}

# Evidence: docs.llmgateway.io/resources/rate-limits - 429 with code rate_limit_exceeded; endpoint at /v1/chat/completions.
llmgateway:
  protocol: openai
  base_url: https://api.llmgateway.io
  error_map:
    rate_limit_exceeded: rate_limit

# Evidence: llmtr.com/docs/billing/ (402 insufficient_balance) + /docs/gateway/chat-completions/ (429 rate_limit_exceeded). Note: bare domain, no api. subdomain.
llmtr:
  protocol: openai
  base_url: https://llmtr.com
  error_map:
    insufficient_balance: billing
    rate_limit_exceeded: rate_limit

# Evidence: docs.meganova.ai/inference-models/text-generation shows curl to https://api.meganova.ai/v1/chat/completions, Bearer.
meganova:
  protocol: openai
  base_url: https://api.meganova.ai
  error_map: {}

# Evidence: docs.mixlayer.com/chat-completions documents https://models.mixlayer.ai/v1/chat/completions, Bearer $MIXLAYER_API_KEY. Note: host is models.mixlayer.AI, docs on mixlayer.COM.
mixlayer:
  protocol: openai
  base_url: https://models.mixlayer.ai
  error_map: {}

# Evidence: moark.ai/docs .../text-comlpetion cites https://api.moark.ai/v1/chat/completions, Bearer access token.
moark:
  protocol: openai
  base_url: https://api.moark.ai
  error_map: {}

# Evidence: modelscope.cn/docs/model-service/API-Inference/intro (verified via official OSS-hosted markdown) - base https://api-inference.modelscope.cn/v1, Bearer access token.
modelscope:
  protocol: openai
  base_url: https://api-inference.modelscope.cn
  error_map: {}

# Evidence: portal.neuralwatt.com/docs/guides/error-handling; billing exhaustion is plain HTTP 402 (no JSON code).
neuralwatt:
  protocol: openai
  base_url: https://api.neuralwatt.com
  error_map:
    concurrent_budget_exceeded: rate_limit
    model_overloaded: rate_limit

# Evidence: opencode.ai/docs/zen/ lists endpoint https://opencode.ai/zen/v1/chat/completions per model; live probe confirms Bearer.
opencode:
  protocol: openai
  base_url: https://opencode.ai/zen
  error_map: {}

# Evidence: opencode.ai/docs/go/ documents https://opencode.ai/zen/go/v1/chat/completions, Bearer.
opencode-go:
  protocol: openai
  base_url: https://opencode.ai/zen/go
  error_map: {}

# Evidence: docs.orcarouter.ai/operations/errors.md (403 insufficient_user_quota); 429s may have NO JSON body - map by HTTP status only.
orcarouter:
  protocol: openai
  base_url: https://api.orcarouter.ai
  error_map:
    insufficient_user_quota: billing

# Evidence: creator.poe.com/docs/external-applications/openai-compatible-api - 402 insufficient_credits, 429 rate_limit_error.
poe:
  protocol: openai
  base_url: https://api.poe.com
  error_map:
    insufficient_credits: billing
    rate_limit_error: rate_limit

# Evidence: qhaigc.net/docs/api-reference/chat.md documents POST https://api.qhaigc.net/v1/chat/completions, Bearer sk-.
qihang-ai:
  protocol: openai
  base_url: https://api.qhaigc.net
  error_map: {}

# Evidence: developer.qiniu.com/aitokenapi/13379 documents base https://api.qnaigc.com/v1, Bearer. Note: domain qnaigc.com, not qiniu.com.
qiniu-ai:
  protocol: openai
  base_url: https://api.qnaigc.com
  error_map: {}

# Evidence: docs.regolo.ai/models/families/completions/ shows https://api.regolo.ai/v1/chat/completions, Bearer.
regolo-ai:
  protocol: openai
  base_url: https://api.regolo.ai
  error_map: {}

# Evidence: docs.requesty.ai/api-reference/endpoint/chat-completions-create - POST https://router.requesty.ai/v1/chat/completions; errors normalized to HTTP statuses (402/412/429).
requesty:
  protocol: openai
  base_url: https://router.requesty.ai
  error_map: {}

# Evidence: docs.stackit.cloud .../use-the-models/ - region-specific (eu01) base, Bearer portal token.
stackit:
  protocol: openai
  base_url: https://api.openai-compat.model-serving.eu01.onstackit.cloud
  error_map: {}

# Evidence: platform.stepfun.ai/docs/en/api-reference/chat/chat-completion-create - Bearer $STEP_API_KEY; China alt api.stepfun.com. Errors are HTTP-status-only (402 billing, 429 rate).
stepfun-ai:
  protocol: openai
  base_url: https://api.stepfun.ai
  error_map: {}

# Evidence: submodel.gitbook.io/docs/instagen/api-reference - base https://llm.submodel.ai/v1, Bearer.
submodel:
  protocol: openai
  base_url: https://llm.submodel.ai
  error_map: {}

# caveat: codes from TokenHub control-plane docs, not data-plane
# Evidence: intl.cloud.tencent.com/document/product/1300/80633 documents the exact endpoint. China region: tokenhub.tencentcloudmaas.com.
tencent-tokenhub:
  protocol: openai
  base_url: https://tokenhub-intl.tencentcloudmaas.com
  error_map:
    LimitExceeded: rate_limit
    RequestLimitExceeded: rate_limit
    RequestLimitExceeded.UinLimitExceeded: rate_limit
    RequestLimitExceeded.IPLimitExceeded: rate_limit

# Evidence: thegrid.ai/docs/api-reference/consumption-api.md - POST https://api.thegrid.ai/v1/chat/completions, Bearer; billing=HTTP 402, rate=429 (no JSON codes for inference).
the-grid-ai:
  protocol: openai
  base_url: https://api.thegrid.ai
  error_map: {}

# Evidence: docs.tinfoil.sh/sdk/direct-api-access.md documents https://inference.tinfoil.sh/v1/chat/completions, Bearer. (Direct HTTPS skips their attestation SDK - acceptable but note it.)
tinfoil:
  protocol: openai
  base_url: https://inference.tinfoil.sh
  error_map: {}

# Evidence: trustedrouter.com/docs - base https://api.trustedrouter.com/v1, Bearer sk-tr- keys, TRUSTEDROUTER_API_KEY.
trustedrouter:
  protocol: openai
  base_url: https://api.trustedrouter.com
  error_map: {}

# Evidence: app.umans.ai/offers/code/docs documents https://api.code.umans.ai/v1/chat/completions, Bearer sk- keys.
umans-ai:
  protocol: openai
  base_url: https://api.code.umans.ai
  error_map: {}

# Evidence: docs.wafer.ai/errors.md enumerates all codes; endpoint https://pass.wafer.ai/v1/chat/completions, Bearer wfr_ keys.
wafer.ai:
  protocol: openai
  base_url: https://pass.wafer.ai
  error_map:
    insufficient_credits: billing
    concurrency_limit_exceeded: rate_limit
    request_quota_exceeded: rate_limit
    rate_limit_exceeded: rate_limit
    backend_at_capacity: rate_limit
    backend_overloaded: rate_limit

# Evidence: xpersona.co/docs curl example POST https://www.xpersona.co/v1/chat/completions, Bearer $XPERSONA_API_KEY. Note: www host, no api. subdomain.
xpersona:
  protocol: openai
  base_url: https://www.xpersona.co
  error_map: {}

# Evidence: docs.zeldoc.ai/use-zeldoc-anywhere/ shows exact curl to https://api.zeldoc.ai/v1/chat/completions, Bearer.
zeldoc:
  protocol: openai
  base_url: https://api.zeldoc.ai
  error_map: {}

# Evidence: zenifra.com/ia documents https://ai.zenifra.com/v1/chat/completions, Bearer $ZENIFRA_AI_API_KEY.
zenifra:
  protocol: openai
  base_url: https://ai.zenifra.com
  error_map: {}

# Evidence: zenmux.ai/docs/guide/advanced/error-codes.html (match on error "type" field); endpoint https://zenmux.ai/api/v1/chat/completions.
zenmux:
  protocol: openai
  base_url: https://zenmux.ai/api
  error_map:
    insufficient_credit: billing
    reject_no_credit: billing
    quote_exceeded: billing
    rate_limit: rate_limit

# -- OpenAI Responses protocol (/v1/responses) ----------------------------------------------------
# The newer stateful OpenAI surface, distinct from chat completions. Bearer auth + a fixed path.
responses:
  protocol: responses
  base_url: https://api.openai.com
  error_map:
    insufficient_quota: billing
    rate_limit_exceeded: rate_limit

# -- Google Gemini protocol (generateContent) -----------------------------------------------------
# The path embeds the model and auth uses x-goog-api-key, both handled by the gemini protocol.
gemini:
  protocol: gemini
  base_url: https://generativelanguage.googleapis.com
  error_map:
    RESOURCE_EXHAUSTED: rate_limit

# Google Vertex AI (Gemini) speaks the SAME gemini protocol, but at a project/location-scoped URL and
# with OAuth (`auth: jwt-bearer` — a signed service-account JWT exchanged for an auto-refreshed bearer)
# instead of an API key. project + location are per-account, so it is not shipped as a live entry -
# copy this template into your config.yaml providers map and fill in your values. `api_key_env` holds
# the service-account JSON (inline) OR a path to the key file. Vertex returns Google's `google.rpc`
# error codes (same `status` tokens as native gemini), so the error_map mirrors gemini's.
#
#   gemini-vertex:
#     protocol: gemini
#     base_url: https://us-central1-aiplatform.googleapis.com
#     path_base: /v1/projects/YOUR_PROJECT/locations/us-central1/publishers/google/models
#     auth: jwt-bearer
#     api_key_env: VERTEX_SA_KEY                       # SA JSON (inline) or a path to the key file
#     error_map:
#       RESOURCE_EXHAUSTED: rate_limit
#
# Claude-on-Vertex is the SAME shape on the `anthropic` protocol: busbar moves the model into the URL
# (`:rawPredict`) and adds the required `anthropic_version` body field automatically. Just point
# path_base at the `anthropic` publisher:
#
#   claude-vertex:
#     protocol: anthropic
#     base_url: https://us-central1-aiplatform.googleapis.com
#     path_base: /v1/projects/YOUR_PROJECT/locations/us-central1/publishers/anthropic/models
#     auth: jwt-bearer
#     api_key_env: VERTEX_SA_KEY
#     error_map: {}

# -- AWS Bedrock protocol (Converse / ConverseStream) ---------------------------------------------
# Auth is native AWS SigV4: set this provider's api_key_env to the AWS credentials as
# `ACCESS_KEY_ID:SECRET_ACCESS_KEY` or `ACCESS_KEY_ID:SECRET_ACCESS_KEY:SESSION_TOKEN`. The region is
# parsed from the base_url host (`bedrock-runtime.<region>.amazonaws.com`) - set it to your region.
# Streaming requests use ConverseStream; busbar decodes the binary
# application/vnd.amazon.eventstream frames and translates them to the caller's protocol.
bedrock:
  protocol: bedrock
  base_url: https://bedrock-runtime.us-east-1.amazonaws.com
  error_map:
    ThrottlingException: rate_limit
    ServiceQuotaExceededException: billing

# -- Cohere v2 Chat protocol (/v2/chat) -----------------------------------------------------------
# Cohere's Command family via the v2 Chat API. Bearer auth; the protocol targets /v2/chat. Errors
# are signalled by HTTP status (classified by the breaker), so error_map is empty.
cohere:
  protocol: cohere
  base_url: https://api.cohere.com
  error_map: {}