API Client

The network layer — how Claude Code talks to the model, handles errors, and supports multiple providers.

Providers

One interface, four backends. The provider is selected by environment variable:

Your Code

Query Engine

Unified Client

API Client Interface

Backends

Anthropic Direct

AWS Bedrock

Google Vertex

Foundry

Env Variable	Provider	Auth Method
ANTHROPIC_API_KEY	Anthropic Direct	API key or OAuth
CLAUDE_CODE_USE_BEDROCK=1	AWS Bedrock	AWS credential chain (env, profile, IAM, SSO)
CLAUDE_CODE_USE_VERTEX=1	Google Vertex AI	GOOGLE_APPLICATION_CREDENTIALS or gcloud
Foundry config	Foundry	Foundry credentials

Request Shape

Every API call sends the system prompt, conversation history, and tool definitions:

TypeScript

{
  model: "claude-sonnet-4-20250514",
  system: assembledSystemPrompt,     // ~10K+ tokens
  messages: conversationHistory,      // grows each turn
  tools: toolDefinitions,             // 44 tool schemas
  stream: true,
  max_tokens: responseLimit,
}

Python

response = client.messages.create(
    model="claude-sonnet-4-20250514",
    system=assembled_system_prompt,
    messages=conversation_history,
    tools=tool_definitions,
    stream=True,
    max_tokens=response_limit,
)

Streaming

Watch how SSE events arrive from the API and map to what you see in the terminal:

POST /v1/messages (stream: true)streaming

SSE Events

Terminal Output

Claude

Text renders progressively as content_block_delta events arrive. Tool calls appear inline as they're detected. Tool input JSON arrives in chunks and is accumulated via a partial JSON parser until complete.

Error Handling & Retries

Error	What Happens	Strategy
429 Rate Limit	Wait and retry	Exponential backoff with jitter (1s → 2s → 4s)
5xx Server Error	Retry up to 3 times	Same backoff strategy
Network Error	Retry with delay	Increasing intervals
401 Unauthorized	Re-authenticate	Prompt for API key or re-launch OAuth
Context Too Large	Trigger compaction	Summarize old messages, retry

Thundering herd prevention

The backoff includes random jitter — each retry waits a slightly different amount of time. This prevents many clients from retrying at the exact same moment after an outage.

Cost Tracking

Every request tracks token usage, feeding into:

Metric	What	Where It Shows
Input tokens	System prompt + messages + tool schemas	Token budget UI
Output tokens	Model response text + tool calls	Token budget UI
Cache tokens	Prompt caching reads/writes	Analytics

Key Source Files

File	Purpose
src/services/api/claude.ts	Main API client wrapper
src/services/api/bootstrap.ts	User profile and session data
src/services/api/	Provider-specific clients (20+ files)
src/utils/auth.ts	Authentication logic