Provider server-side tools

Server-side tools are tools the provider runs for you — Anthropic’s web_search, OpenAI’s code_interpreter, Google’s googleSearch, and so on. Unlike function tools, there’s no round-trip: the model invokes the tool, the provider runs it, and the result comes back in the same response. Opper forwards these tools verbatim to each provider on that provider’s native endpoint. Surcharges (e.g. $10 / 1k web searches) bill as the provider reports them.

Where each tool lives

Server-side tools are endpoint-specific — each provider only accepts its own tools on its own compatibility endpoint. Send them anywhere else and you get a 400 with a hint pointing at the right endpoint.

Provider	Endpoint	Tools
Anthropic	`POST /v3/compat/v1/messages`	`web_search_`, `web_fetch_`, `bash_`, `computer_`, `memory_`, `text_editor_`, `mcp_toolset`, `advisor_*`
OpenAI	`POST /v3/compat/responses`	`web_search`, `file_search`, `code_interpreter`, `image_generation`
Google	`POST /v3/compat/v1beta/models/{model}:generateContent`	`googleSearch`, `codeExecution`, `urlContext`, `googleSearchRetrieval`

Function tools (your own {type:"function",...} definitions) work on every endpoint — the restriction is just for provider-native server tools.

Examples

curl https://api.opper.ai/v3/compat/v1/messages \
  -H "Authorization: Bearer $OPPER_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "anthropic/claude-sonnet-4-6",
    "max_tokens": 1024,
    "messages": [{"role": "user", "content": "What were the top tech headlines today?"}],
    "tools": [{
      "type": "web_search_20250305",
      "name": "web_search",
      "max_uses": 3
    }]
  }'

Responses

Each provider’s native response shape comes back verbatim, so the official Anthropic / OpenAI / Google SDKs parse it without changes:

Anthropic — content[] contains server_tool_use, web_search_tool_result (with the result URLs + encrypted content), and text blocks whose citations[] reference the sources the model actually grounded on.
OpenAI Responses — output[] contains web_search_call items (with the search queries) and message.content[].output_text.annotations[] url_citation entries.
Google — candidates[0].groundingMetadata carries webSearchQueries, groundingChunks[].web (URI + title), groundingSupports, and searchEntryPoint.

Compact responses

Server-tool payloads can get heavy (Anthropic web_search results carry full snippets + per-result encrypted_content). Send X-Opper-Compact-Response: true to strip the bandwidth-heavy fields while preserving citation URLs and titles. Tradeoff: a compact response can’t be replayed to extend a cited multi-turn conversation, because Anthropic citation continuity needs the encrypted_index / encrypted_content that compact mode drops.

Routing through Bedrock and Vertex

When you route Claude through a cloud platform (bedrock/claude-*, vertexai/claude-*), the supported subset shrinks. Opper rejects with a clear 400 when you ask for something the cloud doesn’t carry.

Tool	Direct Anthropic	Vertex AI Claude	Bedrock Claude
`web_search_*`	✓	✓	✗
`web_fetch_*`	✓	✗	✗
`bash_` / `computer_` / `memory_` / `text_editor_`	✓	✓	✓
`code_execution_*`	✓	✗	✗
`mcp_toolset` / `advisor_*`	✓	✗	✗

Source: Anthropic’s Features not supported lists on each platform’s docs page.

Provider docs

For the full parameter set on each tool (filters, citation options, region settings, etc.), consult the provider’s own reference:

Documentation Index

​Where each tool lives

​Examples

​Responses

​Compact responses

​Routing through Bedrock and Vertex

​Provider docs

​See also