LLM Providers

Nine provider slots. Switching is a settings change; the graph is provider-agnostic.

Supported providers

Provider	LangChain wrapper	Default model	Notes
Anthropic	ChatAnthropic	claude-sonnet-4-5	Native SDK
OpenAI	ChatOpenAI	gpt-4o	Native SDK
Google Gemini	ChatGoogleGenerativeAI	gemini-2.5-pro	Native SDK
Groq	ChatGroq	llama-3.1-70b-versatile	Native SDK
xAI Grok	ChatOpenAI	model per grok.model	OpenAI-compatible, base URL x.ai
DeepSeek	ChatOpenAI	deepseek-chat	OpenAI-compatible
Qwen	ChatOpenAI	qwen-max	OpenAI-compatible (DashScope)
GLM (Zhipu AI)	ChatOpenAI	glm-4	OpenAI-compatible
Local	ChatOpenAI	user-set	OpenAI-compatible localhost — Ollama, LM Studio, Docker, OpenRouter

Anything that speaks OpenAI’s API works via the Local slot — set local.baseUrl to point at it.

On xAI Grok: it’s a selectable provider option (generativeAi.option: "XGrok") and key resolution routes it to https://api.x.ai/, but under the hood the canonical PROVIDER_REGISTRY aliases Grok to the Anthropic webview provider rather than giving it a first-class agent-mode entry. Treat the other eight slots as the fully independent ones.

Selection

sequenceDiagram participant Settings participant KM as getAPIKeyAndModel() participant SS as SecretStorage participant Proxy as CredentialProxy participant BC as buildChatModel() Settings->>KM: generativeAi.option = "Anthropic" KM->>SS: Check OS keychain alt Found SS-->>KM: apiKey else Miss KM->>Settings: settings.json fallback end KM->>KM: Reject placeholder keys<br/>("apiKey", "YOUR_API_KEY_HERE") opt Proxy enabled KM->>Proxy: Route via localhost proxy end KM-->>BC: {apiKey, model, baseUrl?, proxyToken?} BC-->>Settings: LangChain ChatModel

Resolution order: SecretStorage (keychain) → settings.json → credential-proxy indirection if enabled. Placeholder strings ("apiKey", "YOUR_API_KEY_HERE", "not-needed") are rejected before the SDK sees them.

Two factory paths

Mode	Factory	Output
Agent	buildChatModel() → LangChain wrapper	Fed into the deepagents pipeline via AgentFactory
Inline completion	CompletionProviderFactory.getProvider()	Returns ICodeCompleter — GroqLLM, QwenLLM, GeminiLLM, LocalLLM

Automatic failover

ProviderFailoverService monitors provider health, classifies errors, and switches to a backup on failure.

Error classes:

Reason	Triggers	Cooldown
auth	HTTP 401, 403	10 min
rate_limit	HTTP 429	1 min
billing	Billing / quota errors	30 min
timeout	Request timeout	30 s
overloaded	HTTP 503, 529	2 min
model_not_found	HTTP 404	1 hr
format	Schema / format errors	—
unknown	Everything else	—

Health state machine: healthy → degraded → down → healthy. A probe recovery attempt fires 30 s before cooldown expires, restoring the provider without waiting out the full timer.

Failover resolution: resolveProvider(primary) checks the primary → walks candidate list respecting cooldowns → returns first with valid creds. The response carries an isFallback flag so callers know a switch happened. Emits provider_switch, health_update, probe_recovery events.

Settings:

Setting	Default	Purpose
codebuddy.failover.enabled	true	Enable automatic failover
codebuddy.failover.providers	[]	Ordered fallback list; empty = auto-detect from configured keys

Cost tracking

CostTrackingService records per-provider token usage against per-model pricing tables (30+ models, separate input / output / cache-read / cache-creation rates). See Cost Tracking.

Provider settings

Full setting list at Settings Reference. Common shape:

Setting	Purpose
generativeAi.option	Active provider — Anthropic, OpenAI, Gemini, Groq, XGrok, Deepseek, Qwen, GLM, Local
{provider}.apiKey	API key (typically anthropic.apiKey, openai.apiKey, google.gemini.apiKeys, etc.)
{provider}.model	Model name for that provider
local.baseUrl	Endpoint for the Local slot