Direct-API Providers (OpenAI-compatible key-loop)

This page covers the OpenAI-compatible direct-key providers registered in Cyberine as loop-*-build build-fleet workers. Each uses LoopAgentAdapter and requires only an API key — no CLI install, no gateway, no Docker.

Excluded from this page (they have dedicated pages): openai, minimax, deepseek, openrouter, moonshot-kimi, anthropic, ollama.

Also excluded: loop-copilot-build (OAuth token, not a static API key) and loop-bedrock-build / loop-vertex-build (SigV4 / ADC auth — see their own sections in the codebase).

Quick reference

All workers: adapter = LoopAgentAdapter, kind = key-loop, auth header = Authorization: Bearer <KEY> (except Azure — see notes below).

Per-provider notes

Google Gemini (loop-gemini-build)

• Endpoint style: Google exposes an OpenAI-compatible shim at generativelanguage.googleapis.com/v1beta/openai. The openai SDK works as-is by overriding baseURL.
• Auth: GEMINI_API_KEY from https://aistudio.google.com/app/apikey — no project or ADC required (distinct from loop-vertex-build which uses Vertex ADC / service account).
• Multimodal: yes (gemini-2.5-pro accepts images).
• Model env: GEMINI_MODEL (default gemini-2.5-pro).
• Gotcha: JSON output can be inconsistent across model versions; tool-call schemas must be precise.

xAI Grok (loop-xai-build)

• Auth: XAI_API_KEY from https://console.x.ai/.
• Multimodal: yes (grok-3 accepts images).
• Model env: XAI_MODEL (default grok-3).
• Cost: frontier tier (cost class 3); reserve for hard / critical tasks.

Groq (loop-groq-build)

• Auth: GROQ_API_KEY from https://console.groq.com/keys.
• Strengths: very low latency, cheap.
• Gotcha: aggressive rate limits on the free tier; quota errors trigger the automatic cooldown and worker re-registration on the next run.
• Model env: GROQ_MODEL (default llama-3.3-70b-versatile).

Mistral (loop-mistral-build)

• Auth: MISTRAL_API_KEY from https://console.mistral.ai/api-keys/.
• Model env: MISTRAL_MODEL (default mistral-large-latest).
• Cost: standard tier (cost class 2).

Together AI (loop-together-build)

• Auth: TOGETHER_API_KEY from https://api.together.xyz/settings/api-keys.
• Model id format: <org>/<model> (e.g. meta-llama/Llama-3.3-70B-Instruct-Turbo).
• Model env: TOGETHER_MODEL (default meta-llama/Llama-3.3-70B-Instruct-Turbo).
• Strengths: cheap, broad open-weights catalog.

Fireworks AI (loop-fireworks-build)

• Auth: FIREWORKS_API_KEY from https://fireworks.ai/account/api-keys.
• Model id format: accounts/fireworks/models/<model> — always the full path.
• Model env: FIREWORKS_MODEL (default accounts/fireworks/models/deepseek-v3).
• Strengths: cheap, fast inference, open-weights catalog.

Qwen / DashScope (loop-qwen-build)

• Auth: QWEN_API_KEY — international (non-mainland) key from https://bailian.console.alibabacloud.com/ (DashScope international account).
• Base URL: international endpoint (dashscope-intl.aliyuncs.com). For mainland China, override with QWEN_BASE_URL=https://dashscope.aliyuncs.com/compatible-mode/v1.
• Model env: QWEN_MODEL (default qwen3-coder-plus).
• Strengths: cheap, strong code generation, long context.

Zhipu GLM (loop-zhipu-build)

• Auth: ZHIPU_API_KEY from https://open.bigmodel.cn/.
• Base URL: https://open.bigmodel.cn/api/paas/v4 (OpenAI-compatible path).
• Model env: ZHIPU_MODEL (default glm-4.6).
• Strengths: cheap, solid instruction following.

Perplexity (loop-perplexity-build)

• Auth: PERPLEXITY_API_KEY from https://www.perplexity.ai/settings/api.
• Base URL: https://api.perplexity.ai (no /v1 suffix — the Cyberine adapter appends the completions path).
• Model env: PERPLEXITY_MODEL (default sonar-pro).
• Strengths: live web search, up-to-date knowledge, inline citations.
• Gotcha: Sonar models are search-optimised; they are less suited to pure code generation than the other workers in this fleet. Use for research/Q&A steps.
• Note: PERPLEXITY_API_KEY also backs the web_search tool used by all loop workers. Setting it registers both the search tool AND this build worker.

Cerebras (loop-cerebras-build)

• Auth: CEREBRAS_API_KEY from https://cloud.cerebras.ai/.
• Model env: CEREBRAS_MODEL (default llama-3.3-70b).
• Strengths: ultra-fast inference (wafer-scale chip), very cheap.
• Gotcha: capability ceiling — the current Llama 3.3 70B default is strong for its price point but weaker than frontier models on complex reasoning tasks.

NVIDIA NIM (loop-nvidia-build)

• Auth: NVIDIA_API_KEY from https://build.nvidia.com/explore/discover.
• Base URL: https://integrate.api.nvidia.com/v1 (OpenAI-compatible).
• Model env: NVIDIA_MODEL (default meta/llama-3.3-70b-instruct).
• Model id format: <org>/<model> (e.g. meta/llama-3.3-70b-instruct).
• Strengths: broad model catalog (Llama, Mixtral, Nemotron, etc.), optimised NIM containers.

Cohere (loop-cohere-build)

• Auth: COHERE_API_KEY from https://dashboard.cohere.com/api-keys.
• Base URL: https://api.cohere.ai/compatibility/v1 — Cohere's OpenAI-compatible shim (distinct from their native /v1 endpoint).
• Model env: COHERE_MODEL (default command-r-plus).
• Strengths: retrieval-augmented generation, solid instruction following.

Azure OpenAI (loop-azure-build)

• Auth: AZURE_OPENAI_API_KEY — resource-scoped, not a personal key.
• Endpoint: AZURE_OPENAI_ENDPOINT — the full deployment URL for your Azure resource, e.g. https://<resource>.openai.azure.com/openai/deployments/<deployment>/.
• Deployment name: AZURE_OPENAI_DEPLOYMENT (default gpt-4o) — the deployment id inside your resource, not the model id.
• API version: AZURE_OPENAI_API_VERSION (default 2024-12-01-preview) — appended as an ?api-version= query parameter on every request.
• Multimodal: yes (gpt-4o default accepts images).
• Gotcha: both AZURE_OPENAI_API_KEY and AZURE_OPENAI_ENDPOINT must be set for the worker to register; missing either skips the adapter with a note.
• Cost: frontier tier (cost class 3); suited for enterprise compliance scenarios where data must not leave Azure.

Xiaomi MiMo (loop-mimo-build)

• Auth: XIAOMI_API_KEY — funded direct key from Xiaomi's AI platform.
• Base URL: https://api.xiaomimimo.com/v1 (OpenAI-compatible). Regional variants exist (e.g. token-plan-cn.xiaomimimo.com, token-plan-ams.xiaomimimo.com, token-plan-sgp.xiaomimimo.com) — override with XIAOMI_BASE_URL.
• Model env: XIAOMI_MODEL (default mimo-v2.5-pro).
• Strengths: cheap, long context, surgical edits.
• Note: also registered as loop-mimo-direct in the default (read-only) adapter set when XIAOMI_API_KEY is set; loop-mimo-build is the implementer (write-capable) build worker.
• Model catalog (T651a, 2026-08-27, live-probed with a real XIAOMI_API_KEY): the catalog used to also list mimo-v2-pro / mimo-v2-flash / mimo-v2-omni behind a CONTRADICTED deprecation signal (one source said deprecated 2026-06-30, two others didn't). Resolved by GET /v1/models (returns exactly 7 live ids: mimo-v2.5, mimo-v2.5-asr, mimo-v2.5-pro, mimo-v2.5-pro-ultraspeed, mimo-v2.5-tts, mimo-v2.5-tts-voiceclone, mimo-v2.5-tts-voicedesign — none of the three non-2.5 chat ids appear) plus a 1-token POST /v1/chat/completions per id: mimo-v2.5-pro and mimo-v2.5 both 200 with a real completion; mimo-v2-pro / mimo-v2-flash / mimo-v2-omni all 400 {"error":{"code":"400","message":"Unsupported model mimo-v2-<id>"}}. The three non-2.5 ids were removed from modelCatalog.ts / modelPricing.ts (and the isMultimodalModel() regex in workerProfiles.ts that matched mimo-v2-omni). mimo-v2.5-pro-ultraspeed (untested faster variant) and the -asr/-tts* ids (different modality) were seen live but NOT added — a future sweep should probe -pro-ultraspeed for the chat catalog.

Usage

Enable one or more workers in CYBERINE_BUILD_FLEET:

CYBERINE_BUILD_FLEET="loop-gemini-build,loop-groq-build" npm run cli -- "<task>"

Override the model for a specific worker:

GEMINI_MODEL=gemini-2.0-flash-001 CYBERINE_BUILD_FLEET="loop-gemini-build" npm run cli -- "<task>"

All workers inherit the global LOOP_MAX_TOOL_ROUNDS (default 25) and LOOP_MAX_TOKENS (default 16000) limits.