Five of eight DuckDuckGo MCP servers failed calls on the same twelve back-to-back queries

A call that comes back with nothing in it looks, to an agent, like a tool call that worked. The harness counted it as a failure, and the post says that third rule "did most of the work" ^[4]. The medians make the same trap worse, because they only count calls that returned something, so the three worst servers look like the quick ones. "OEvortex failing in 1.53 s is still a failure," the write-up says ^[5].

The fastest finisher also ships a default I would change before putting it in front of a coding agent. free-search-mcp answered twelve of twelve in all three passes at a 1.39 s median, out of a 69-star MIT project last pushed in August 2026 ^[6]. It returns markdown, and JSON only when the call asks for it, which means passing `query`, `engines`, `format` and `use_cache` every time ^[7]. `use_cache` defaults to true, and will hand back an old answer if you leave it alone ^[7]. That is a sane default for a person browsing and the wrong one for an agent checking whether a library API moved last week.

Two registries carry the same package name. `uvx duckduckgo-mcp-server` gets nickclyde's server from PyPI; `npx -y duckduckgo-mcp-server` gets zhsama's unrelated package, which scored zero of twelve in all three passes ^[11]. According to the post, Google's AI Overview for that package name currently recommends the npm version ^[12].

nickclyde's is the popular one: 1,460 stars, MIT, the base of the `mcp/duckduckgo` Docker image with over half a million pulls, and twelve of twelve in every pass at a 2.37 s median ^[8]. A fourth run from a fresh clone answered seven and refused five, each refusal citing DuckDuckGo's bot detection and suggesting the optional browser backend that impersonates Chrome's TLS ^[9]. For any of the twelve-of-twelve scores to transfer to your setup, DuckDuckGo would have to treat your egress address and TLS fingerprint the way it treated one machine on one IP on 2026-09-15 ^[3].

HasData, which employs the author and runs one of the eight servers, finished twelve of twelve in all three passes at a 1.96 s median, third on speed in the first pass and fourth in the two after, behind free servers ^[13][14]. Its payload is the heaviest: ten results carrying position, title, link, displayedLink, source, snippet and snippetHighlightedWords, plus DuckDuckGo's assist answer and a pagination token ^[15]. The payload runs 10.0 KB per call against nickclyde's 3.6 KB ^[16], so a twelve-query dig costs about 76.8 KB more context ^[19]. Whether the weight is worth it depends on how reliably your model parses `Found N search results:` and a numbered list, which is what nickclyde's server returns, with no pagination ^[10].

Cold config to first answer was 2.3 s on the hosted endpoint, 8.3 s for free-search and 8.8 s for nickclyde, while open-webSearch and OEvortex took 37 to 48 s resolving and building packages ^[17]. You pay that once per config change, not per call.

The gap between the two fastest finishers is 0.98 s a call, or roughly 11.8 s across a twelve-query run ^[21]. Five of the eight never got that far ^[20]. The harness is public ^[18].