Providers

OpenCorvus LLM provider configuration and model selection.

OpenCorvus’s LLM layer sits on top of Vercel AI SDK, using @ai-sdk/* packages for each provider.

Built-in providers

From packages/opencorvus/package.json:

Provider	Package
Anthropic	`@ai-sdk/anthropic`
OpenAI / compatible	`@ai-sdk/openai` + `@ai-sdk/openai-compatible`
Google	`@ai-sdk/google` + `@ai-sdk/google-vertex`
Amazon Bedrock	`@ai-sdk/amazon-bedrock`
Azure OpenAI	`@ai-sdk/azure`
Cerebras / Cohere / DeepInfra / Groq / Mistral / Perplexity / TogetherAI / Vercel / xAI	`@ai-sdk/*`
GitLab AI	`@gitlab/gitlab-ai-provider`
OpenRouter	`@openrouter/ai-sdk-provider`
Vercel Gateway	`ai-gateway-provider`

China-region aliases (declared in channel-runtime/.env.example):

Alias	Underlying
`alibaba-cn`	DashScope
`alibaba-coding-plan-cn`	DashScope Coding Plan
`moonshotai-cn`	Moonshot
`deepseek`	DeepSeek

Configuration

A. Environment variables

Each provider’s canonical key name works out of the box:

export ANTHROPIC_API_KEY=sk-ant-...
export OPENAI_API_KEY=sk-...
export GOOGLE_GENERATIVE_AI_API_KEY=...
export DEEPSEEK_API_KEY=...
export OPENROUTER_API_KEY=...

# Alibaba DashScope
export DASHSCOPE_API_KEY=sk-...                 # alibaba-cn and shared DashScope key
export ALIBABA_CODING_PLAN_API_KEY=sk-sp-...    # alibaba-coding-plan-cn

B. Custom provider (OpenAI-compatible)

In opencorvus.jsonc:

{
  "provider": {
    "myhub": {
      "api": "https://your-gateway/v1",
      "env": ["MYHUB_API_KEY"],
      "models": {
        "custom-1": { "name": "Custom 1", "tool_call": true },
      },
    },
  },
  "model": "myhub/custom-1",
}

env declares required env names; missing keys fail loud — no fallback. Provider base URLs come from the model database; key prefixes do not override provider routing.

Model selection

Model selection has explicit identity layers:

Scenario	Config
Project default	`model`
Fixed Primary/Helper/Host identity	`agent.<fixed-id>.model`
Worker template	`runtime_templates.<base-role>.model`
Exact projected worker	`expert_squads.<squad-id>.agents.<agent-id>.runtime.model`
Vision / screenshot understanding	`OPENCORVUS_VISION_MODEL`

Splitting planning from execution lets you plan with a strong/slow model and execute with a fast/cheaper one:

{
  "model": "openai/gpt-5.5",
  "agent": {
    "coding": { "model": "anthropic/claude-sonnet-4-6" },
  },
  "runtime_templates": {
    "architect": { "model": "anthropic/claude-sonnet-4-6" },
  },
}

Reasoning-model constraints

For Claude reasoning, qwq, o1, and similar:

Must use streamText — generateText hangs while reasoning tokens stream.
Must use toolChoice: "auto", not "required".
Prompt caching matters (5-minute TTL for Anthropic reasoning models). Reuse system prompts across turns.

Excluding a provider

To avoid using a provider, simply don’t configure its env — OpenCorvus removes keyless providers from the available list. Do not rely on heuristic filtering.