Anthropic Tool Calling

Latest Audit Result

The Anthropic Tool Calling guide currently covers only one Exa capability: `search` with `type='auto'` and highlights. That Exa search call shape is broadly consistent with current docs, but the guide is outdated as a Claude tool-use example because it executes the tool locally and then appends plain assistant/user text instead of preserving the assistant `tool_use` message and returning a user `tool_result` block with the matching `tool_use_id`. The Exa changelog deprecates `/research`, `resolvedSearchType`, `highlightScores`, and `startCrawlDate`/`endCrawlDate`; this guide does not use those deprecated features. Several previously flagged feature gaps remain relevant as nice-to-have coverage, except `/research` should not be added because it is sunset in favor of `/search` with `type: "deep-reasoning"`.

Missing Features (9)

• The guide's Anthropic tool-use code is outdated/incorrect: it does not return `tool_result` content blocks to Claude and does not include the original `tool_use` assistant message in the conversation. Current Anthropic docs state client tools return `stop_reason: "tool_use"`, the app executes the tool, and sends back a `tool_result`.
• The guide imports `load_dotenv` but the install command omits `python-dotenv`, so the full code example may fail in a clean environment with `ModuleNotFoundError: dotenv`.
• The introductory JSON snippet is labeled `json` but contains `#` comments, which is invalid JSON.
• The full-code comment links to `https://docs.exa.ai/reference/python-sdk-specification#search-method`, but the fetched docs redirect behavior did not surface that page; current Exa docs expose `/reference/search` and SDK examples there instead.
• DEPRECATED: The guide does not use deprecated Exa features (`/research`, `resolvedSearchType`, `highlightScores`, `startCrawlDate`, or `endCrawlDate`), so no deprecated usage was found in this integration.
• Update the Anthropic tool-use loop to follow the current Claude API pattern: after `stop_reason: "tool_use"`, append Claude's assistant message containing the `tool_use` block, execute Exa locally, then send a user message containing a `tool_result` block with the original `tool_use.id` before asking Claude for the final answer.
• Improve the Exa tool schema for the existing `exa_search` tool by exposing at least `numResults` and optional search/content controls already supported by `/search`, instead of accepting only `query`.
• Use current content-control guidance in the existing highlights example: `contents={"highlights": True}` is valid, but docs now prefer showing `highlights` as `true` or an object with `maxCharacters` when tuning length; avoid deprecated `numSentences`/`highlightsPerUrl`.
• The live docs version of the guide differs from the local MDX on the Exa call: live docs say `search_and_contents(..., highlights=True)`, while local MDX says `search(..., contents={'highlights': True})`. For a docs-only baseline, align the published guide and source to one current pattern.