Anthropic Provider on Bifrost

Anthropic provider summary

Bifrost supports Anthropic models through OpenAI-compatible APIs and Anthropic-compatible passthrough routes.

Supported OpenAI parameters

Quick list of OpenAI-style parameters accepted while targeting Anthropic through Bifrost.

[
  "stream",
  "stop",
  "temperature",
  "top_p",
  "max_tokens",
  "max_completion_tokens",
  "tools",
  "tool_choice",
  "response_format",
  "user",
  "reasoning",
  "top_k"
]

Use as a Claude Code backend

Claude Code can use Bifrost as an Anthropic-compatible backend by pointing ANTHROPIC_BASE_URL to Bifrost's Anthropic endpoint. This is a common setup for routing Claude traffic through governance, logs, and fallback.

# Point Claude Code to Bifrost's Anthropic-compatible endpoint
export ANTHROPIC_BASE_URL="http://localhost:8080/anthropic"

# Use your Bifrost virtual key (or provider key depending on setup)
export ANTHROPIC_API_KEY="vk_your_key"

# Launch Claude Code
claude

For full walkthrough and troubleshooting, see Claude Code gateway setup.

Supported operations

Anthropic has different payload structures than OpenAI. Bifrost translates requests and responses so your existing OpenAI-style clients can still work. Embeddings, Speech, Transcriptions, and Image Generation are not supported by the upstream Anthropic API. These return UnsupportedOperationError.

Beta headers

Bifrost automatically manages Anthropic beta headers by detecting required headers from request features and injecting them. Headers are validated per provider to prevent unsupported headers from reaching the upstream API. Passthrough headers are forwarded when set manually via the anthropic-beta request header.

API reference by operation

Gateway paths and upstream Anthropic endpoints, aligned with the Bifrost Anthropic provider docs.

curl -X POST http://localhost:8080/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "anthropic/claude-3-5-sonnet",
    "messages": [{"role": "user", "content": "Hello"}],
    "top_k": 40
  }'

curl -X POST http://localhost:8080/v1/responses \
  -H "Content-Type: application/json" \
  -d '{
    "model": "anthropic/claude-3-5-sonnet",
    "input": "Hello, how are you?",
    "top_k": 40
  }'

curl -X POST http://localhost:8080/v1/files \
  -F "file=@document.pdf"

curl http://localhost:8080/v1/models

Prompt caching (Anthropic)

Add cache directives to system messages, user messages, and tool definitions. Bifrost also forwards a top-level cache_control object on /anthropic/v1/messages requests unchanged.

Caching example

curl -X POST http://localhost:8080/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "anthropic/claude-3-5-sonnet",
    "messages": [
      {
        "role": "user",
        "content": [
          {
            "type": "text",
            "text": "This is cached context",
            "cache_control": {"type": "ephemeral"}
          }
        ]
      }
    ],
    "system": [
      {
        "type": "text",
        "text": "You are a helpful assistant",
        "cache_control": {"type": "ephemeral"}
      }
    ]
  }'

Reasoning / thinking

reasoning.effort maps to thinking.type (always "enabled"). reasoning.max_tokens maps to thinking.budget_tokens. Minimum budget is 1024 tokens; -1 is converted to 1024 automatically.

// Request
{"reasoning": {"effort": "high", "max_tokens": 2048}}

// Anthropic conversion
{"thinking": {"type": "enabled", "budget_tokens": 2048}}

Dropped parameters

The following OpenAI-style parameters are silently ignored when targeting Anthropic through Bifrost:

[
  "frequency_penalty",
  "presence_penalty",
  "logit_bias",
  "logprobs",
  "top_logprobs",
  "seed",
  "parallel_tool_calls",
  "service_tier"
]

Supported Claude models

Use the provider prefix anthropic/ in Bifrost model routes for deterministic provider targeting.

Core request mapping

Bifrost normalizes OpenAI-format input to Anthropic-format fields. This is the key layer that enables compatibility without changing client code.

Implementation caveats

Response conversion

Key field mappings on the way back to OpenAI-compatible responses:

• stop_reason → finish_reason (end_turn/stop_sequence → stop, max_tokens → length, tool_use → tool_calls)
• Cache token counts roll into prompt_tokens with breakdown in prompt_tokens_details
• thinking blocks → reasoning_details
• Streaming: message_start → content_block_delta → message_stop

Anthropic passthrough

Send native Anthropic Messages API payloads to /anthropic/v1/messages when you need full control over Anthropic request shapes.

curl -X POST http://localhost:8080/anthropic/v1/messages \
  -H "Content-Type: application/json" \
  -H "x-bf-vk: vk_your_key" \
  -d '{
    "model": "claude-sonnet-4-5-20250929",
    "max_tokens": 1024,
    "messages": [{"role": "user", "content": "What changed between versions?"}]
  }'