Tavily

tavily-restrestprimary

78% coverage

Offload Valuejudged · weight 15 · Oct 5, 2026

Tavily Search retrieves live web content the model cannot generate from weights: real-time indexed results, publish-date estimates, domain/country/language filtering, per-source semantically relevant chunk extraction, cleaned/parsed HTML, image extraction with descriptions, and an optional LLM-generated synthesized answer. This is genuine external data acquisition plus retrieval ranking and content distillation, all of which would be impossible or extremely costly for an agent to simulate through inference (and a naive agent would otherwise fetch and read whole pages, burning huge context). It falls short of the top band only because it performs no irreversible real-world side effect (no funds, signing, or filing) and because fetch-plus-summarize is approximable by a resourceful agent with raw HTTP access, albeit at much higher token cost.

• https://docs.tavily.com/documentation/api-reference/endpoint/search

  Chunks are short content snippets (maximum 500 characters each) pulled directly from the source. Use chunks_per_source to define the maximum number of relevant chunks returned per source and to control the content length.

• https://docs.tavily.com/documentation/api-reference/endpoint/search

  Highest relevance with increased latency. Best for detailed, high-precision queries. Returns multiple semantically relevant snippets per URL

• https://docs.tavily.com/documentation/api-reference/endpoint/search

  advanced : Highest relevance with increased latency. Best for detailed, high-precision queries. Returns multiple semantically relevant snippets per URL

Interaction Costmeasured · weight 15 · Oct 5, 2026

Median over 2 successful run(s): 183629 tokens, 9.5 calls. Ratio to the best in "web-search" (3 surfaces measured, v0.2 scale). Cold, without documentation: 1/1 succeeded.

Error Recoverabilityprobed · weight 15 · Oct 5, 2026

unknown_endpoint → HTTP 404, clarity 25/100 (field named: no, allowed values: no, RFC 9457: no) · wrong_type → HTTP 422, clarity 75/100 (field named: yes, allowed values: yes, RFC 9457: no) · hallucinated_enum → HTTP 400, clarity 75/100 (field named: yes, allowed values: yes, RFC 9457: no) · unknown_field → HTTP 200, clarity 40/100 (field named: no, allowed values: no, RFC 9457: no)

• https://api.tavily.com/search

  HTTP 401: { "detail": { "error": "Unauthorized: missing or invalid API key." } }

• https://api.tavily.com/nonexistent_resource_xyz

  HTTP 404:

• https://api.tavily.com/search

  HTTP 422: {"detail":[{"type":"literal_error","loc":["body","max_results","literal['auto']"],"msg":"Input should be 'auto'","input":"ten","ctx":{"expected":"'auto'"}},{"type":"int_parsing","loc":["body","max_res

• https://api.tavily.com/search

  HTTP 400: {"detail":{"error":"Invalid search depth. Must be 'ultra-fast', 'fast', 'basic' or 'advanced'."}}

Doc Legibilitymeasured · weight 12 · Oct 5, 2026

text/markup ratio 4.8% (weight 30) · renders without JavaScript: yes (weight 25) · machine-readable spec: none found (weight 25) · llms.txt: https://docs.tavily.com/llms.txt (weight 10) · 12 runnable example(s) (weight 10). Weighted average over 100 points of available sub-criteria.

• https://docs.tavily.com/documentation/api-reference/endpoint/search
• https://docs.tavily.com/llms.txt

Auth Frictionmeasured · weight 10 · Oct 5, 2026

Authorization: Bearer <key>, verified by a successful call. One header, no ceremony.

• https://api.tavily.com/search

  HTTP 401: { "detail": { "error": "Unauthorized: missing or invalid API key." } }

• https://api.tavily.com/nonexistent_resource_xyz

  HTTP 404:

• https://api.tavily.com/search

  HTTP 422: {"detail":[{"type":"literal_error","loc":["body","max_results","literal['auto']"],"msg":"Input should be 'auto'","input":"ten","ctx":{"expected":"'auto'"}},{"type":"int_parsing","loc":["body","max_res

• https://api.tavily.com/search

  HTTP 400: {"detail":{"error":"Invalid search depth. Must be 'ultra-fast', 'fast', 'basic' or 'advanced'."}}

Safety & Reversibilityprobed · weight 10

If I retry after a timeout, do I charge the customer twice?

Payload Efficiencyprobed · weight 8

Can I ask for only the fields I need, or do I have to swallow 200 attributes to extract one?

Task Atomicitymeasured · weight 6 · Oct 5, 2026

Median over 2 successful run(s): 183629 tokens, 9.5 calls. Ratio to the best in "web-search" (3 surfaces measured, v0.2 scale). Cold, without documentation: 1/1 succeeded.

Prior Knowledge Coveragemeasured · weight 5 · Oct 5, 2026

Median over 2 successful run(s): 183629 tokens, 9.5 calls. Ratio to the best in "web-search" (3 surfaces measured, v0.2 scale). Cold, without documentation: 1/1 succeeded.

Async Compatibilityprobed · weight 2

If the operation is asynchronous, can I get the result without hosting a web server?

Tool-Native (MCP)binary · weight 2

Does the vendor publish an official, maintained MCP server?