Best Web Search MCP for Codex, 2026
Best Web Search MCP for Codex: Exa deep leads 9 search APIs at 83.0% task completion on 100 coding tickets, independently benchmarked in 2026.
What this page compares. This page ranks measured results from 2 boards: coding-agent documentation tickets (search and page fetch) and coding-agent documentation tickets (search only). The benchmark agent runs on GPT-5.6, not inside this editor. The ranking measures the search API behind each MCP server; MCP transport and the client's tool selection were not tested.
How to read the result. The table ranks every complete run by its score on the same questions. Latency and cost are shown beside it and not blended into the score.
Coding-agent documentation tickets, search and page fetch, ranked by task completion
Every row ran the same 100 questions. Provider names link to the official docs for the configuration measured.
| Rank | Provider | Task completion | Mean search time | Median task time | Median cost / task | Median task tokens | Completion / 1k tokens | List price |
|---|---|---|---|---|---|---|---|---|
| 1 | Exa deeptype=deep | 83.0% | 4.0 s | 36.8 s | $0.13 | 23,660 | 3.51 | $12.00 / 1k queries |
| 2 | Exa autotype=auto | 81.7% | 1.2 s | 22.7 s | $0.11 | 27,433 | 2.98 | $7.00 / 1k queries |
| 3 | TinyFishtinyfish-search | 79.0% | 1.3 s | 24.3 s | $0.050 | 12,844 | 6.15 | $0 / 1k queries |
| 4 | Perplexitysearch_context_size=high | 77.7% | 991 ms | 21.5 s | $0.083 | 20,062 | 3.87 | $5.00 / 1k queries |
| 5 | Parallel advancedmode=advanced | 77.0% | 3.1 s | 33.4 s | $0.096 | 27,092 | 2.84 | $5.00 / 1k queries |
| 6 | Firecrawlfirecrawl-search | 76.0% | 2.8 s | 34.4 s | $0.082 | 17,379 | 4.37 | $5.00 / 1k queries |
| 7 | Parallel basicmode=basic | 76.0% | 1.6 s | 30.0 s | $0.11 | 32,809 | 2.32 | $5.00 / 1k queries |
| 8 | Nimblesearch_depth=lite · full_content=false · focus=general | 60.3% | 1.9 s | 40.0 s | — | 16,828 | 3.59 | $1.10 / 1k queries |
| 9 | Tavily advancedsearch_depth=advanced | 60.0% | 3.4 s | 46.4 s | $0.15 | 26,269 | 2.28 | $16.00 / 1k queries |
| 10 | Tavily basicsearch_depth=basic | 59.0% | 1.5 s | 32.4 s | $0.11 | 27,405 | 2.15 | $8.00 / 1k queries |
| 11 | Youextraction_mode=highlights | 55.0% | 638 ms | 27.3 s | $0.12 | 42,806 | 1.28 | $5.00 / 1k queries |
| 12 | Youextraction_mode=highlights · knowledge=core | 54.0% | 678 ms | 27.8 s | $0.12 | 37,453 | 1.44 | $5.00 / 1k queries |
| 13 | Linkup standarddepth=standard | 48.3% | 2.0 s | 48.2 s | $0.15 | 57,791 | 0.84 | $5.00 / 1k queries |
| 14 | Nimblesearch_depth=standard · full_content=false · focus=general | 45.0% | 905 ms | 44.2 s | — | 28,586 | 1.57 | $5.00 / 1k queries |
Coding-agent documentation tickets, search only, ranked by task completion
| Rank | Provider | Task completion | Mean search time | Median task time | Median cost / task | Median task tokens | Completion / 1k tokens | List price |
|---|---|---|---|---|---|---|---|---|
| 1 | Perplexitysearch_context_size=low | 77.3% | 957 ms | 18.3 s | $0.059 | 8,765 | 8.82 | $5.00 / 1k queries |
| 2 | Firecrawlfirecrawl-search | 70.3% | 2.9 s | 28.0 s | $0.059 | 7,456 | 9.43 | $5.00 / 1k queries |
| 3 | Parallel fastmode=fast | 66.7% | 953 ms | 22.1 s | $0.054 | 12,460 | 5.35 | $1.00 / 1k queries |
| 4 | Exa fasttype=fast | 66.3% | 626 ms | 20.0 s | $0.10 | 22,344 | 2.97 | $7.00 / 1k queries |
| 5 | Parallel turbomode=turbo | 64.7% | 333 ms | 18.7 s | $0.059 | 14,130 | 4.58 | $1.00 / 1k queries |
| 6 | Exa instanttype=instant | 61.3% | 447 ms | 20.6 s | $0.10 | 22,423 | 2.74 | $7.00 / 1k queries |
| 7 | TinyFishtinyfish-search | 59.3% | 2.1 s | 27.8 s | $0.035 | 7,469 | 7.94 | $0 / 1k queries |
| 8 | Tavily basicsearch_depth=basic | 51.0% | 1.7 s | 26.4 s | $0.097 | 16,299 | 3.13 | $8.00 / 1k queries |
| 9 | Nimblesearch_depth=lite · full_content=false · focus=general | 46.7% | 3.2 s | 32.7 s | $0.043 | 8,713 | 5.36 | $1.10 / 1k queries |
| 10 | Linkup fastdepth=fast | 43.3% | 1.4 s | 26.8 s | $0.10 | 24,057 | 1.80 | $5.00 / 1k queries |
| 11 | Nimblesearch_depth=standard · full_content=false · focus=general | 42.3% | 878 ms | 25.3 s | $0.079 | 14,859 | 2.85 | $5.00 / 1k queries |
| 12 | Youextraction_mode=highlights | 41.3% | 596 ms | 23.9 s | $0.097 | 20,333 | 2.03 | $5.00 / 1k queries |
| 13 | Youextraction_mode=highlights · knowledge=core | 38.3% | 677 ms | 21.1 s | $0.096 | 20,786 | 1.84 | $5.00 / 1k queries |
| 14 | Bravellm-context | 38.0% | 523 ms | 19.9 s | $0.082 | 14,103 | 2.69 | $5.00 / 1k queries |
Official MCP servers for the tested search APIs
Scores and timings measure the underlying search APIs in our developer-agent benchmark. MCP connection options and tools are verified against vendor documentation; MCP transport overhead and client tool selection require a separate end-to-end test.
- Perplexity.
https://api.perplexity.ai/mcp. Tools: perplexity_search, perplexity_ask, perplexity_research, perplexity_reason. Use perplexity_search for ranked web results. The answer and research tools add a separate answering workflow. Docs ↗ - Exa.
https://mcp.exa.ai/mcp. Tools: web_search_exa, web_fetch_exa, web_search_advanced_exa, agent_run. Search for relevant web content and scrape known URLs. Opt into advanced filters or authenticated agent research when needed. Docs ↗ - Firecrawl.
https://mcp.firecrawl.dev/v2/mcp. Tools: firecrawl_search, firecrawl_scrape, firecrawl_parse, firecrawl_crawl. Search, scrape and parse content through one integration. Authenticated connections expose additional tools according to the account plan. Docs ↗ - Parallel.
https://search.parallel.ai/mcp. Tools: web_search, web_fetch. Search from an objective and keyword queries, then read relevant excerpts from selected URLs. Docs ↗ - Tavily.
https://mcp.tavily.com/mcp. Tools: tavily-search, tavily-extract. Search current sources with filters, then extract page content for the agent's answer. Docs ↗ - Brave Search.
npx -y @brave/brave-search-mcp-server --transport stdio. Tools: brave_web_search, brave_local_search, brave_news_search, brave_image_search, brave_video_search. Search the web or a dedicated news, image, video or local-results surface using Brave's search index. Docs ↗ - TinyFish.
https://agent.tinyfish.ai/mcp. Tools: search, fetch_content, run_web_automation. Get ranked search results, scrape page content or add browser automation when the workflow needs interactions. Docs ↗ - Linkup.
https://mcp.linkup.so/mcp. Tools: linkup-search, linkup-research, linkup-get-research. Search with domain and date controls, or start a longer research task and retrieve its result. Docs ↗ - Nimble.
https://mcp.nimbleway.com/mcp. Tools: nimble_search, nimble_agents_list. Search live web data and discover available agents. The MCP integration also covers extraction, mapping and crawling. Docs ↗ - You.com.
https://api.you.com/mcp. Tools: you-search, you-contents, you-research. Search web and news results, extract page content, or request a source-backed research answer. Docs ↗
How the ranking was measured
- The same coding agent worked the same documentation tickets with one search API at a time, three independent runs per ticket.
- A ticket passes only when the patch checks out and cites a source the agent actually retrieved in that run.
- Search-only rows return results; search-and-fetch rows can also read the page.
- Search time, total task time, task tokens and the median cost per task (model plus API) are reported beside completion.
- The runner and published data are open in openbenchmarks-labs/web-search-for-coding-agents ↗ and openbenchmarks/OB-Code-Websearch ↗. The full methodology is on the Web Search for Coding Agents Benchmark →
Questions answered by this comparison
Which provider leads best Web Search MCP for Codex?
Exa deep leads this table at 83.0% task completion, on 100 coding tickets.
How was this measured?
The same coding agent worked the same documentation tickets with one search API at a time, three independent runs per ticket. A ticket passes only when the patch checks out and cites a source the agent actually retrieved in that run. Search-only rows return results; search-and-fetch rows can also read the page. Search time, total task time, task tokens and the median cost per task (model plus API) are reported beside completion.
How should speed and cost be read beside the score?
Exa deep recorded 36.8 s median task time and $0.13 median cost per task. The most accurate configuration is rarely the fastest or the cheapest, so each is reported in its own column.
Do vendors pay to be ranked?
No. Inclusion and rank do not depend on vendor payments. Every API used the same questions, answering model and scoring.
What does this page not test?
This page ranks measured results from 2 boards: coding-agent documentation tickets (search and page fetch) and coding-agent documentation tickets (search only). The benchmark agent runs on GPT-5.6, not inside this editor. The ranking measures the search API behind each MCP server; MCP transport and the client's tool selection were not tested.