One query in words. The right platforms, merged.

Try the plan for 0 credits#

Add dry_run=true and nothing runs and nothing is charged. You get the plan, each search that would run with its price, and your balance. Recorded on the dev API (shortened), for a query with two platform hints:

curl "https://api.onekey.dev/v1/search/universal?q=rust%20async%20runtime%20on%20hackernews%20and%20devto&dry_run=true" \
  -H "x-api-key: YOUR_API_KEY"

{
  "success": true,
  "data": {
    "plan": {
      "method": "deterministic",
      "cleaned": "rust async runtime",
      "selection": "hints",
      "platforms": ["hackernews", "devto"],
      "subqueries": { "hackernews": "rust async runtime", "devto": "rust" },
      "adapted": { "devto": "first keyword only" }
    },
    "legs": [
      { "platform": "hackernews", "endpoint": "hackernews.search", "params": { "q": "rust async runtime" }, "credits": 1.0, "worst_case_credits": 1.0 },
      { "platform": "devto", "endpoint": "devto.search", "params": { "q": "rust" }, "credits": 5.0, "worst_case_credits": 5.0 }
    ],
    "credits": { "legs_worst_case": 6.0, "legs_list_price": 6.0, "ai": null },
    "ai_enabled": false
  },
  "metadata": { "recipe": "search_universal", "dry_run": true, "legs": 2, "credits_charged": 0.0 }
}

Read it: the words "on hackernews and devto" chose the platforms and were removed from the query. Hacker News got the whole query. The DEV Community search accepts only one tag-style word, so it got the first keyword, and the plan says so in adapted. The whole call would cost at most 10 credits. Remove dry_run to run it.

What runs today, and what does not#

With AI not enabled, a request that asks for relevance or themes and keeps the default strict=true is refused with AI_UNAVAILABLE before anything is fetched, so it costs nothing. With strict=false you get the fused list back unjudged. plan=ai falls back to the deterministic plan and says why in plan.ai. Details and the pre-release test figures are on AI pricing. This page makes no claim about the quality of AI results, because there are none to show.

Platforms it can search#

19 platforms have a search endpoint today. Without hints, the planner picks among them in a fixed order: stable before beta, then cheaper first, then by id, capped at max_platforms (default 6, at most 12). Name platforms with platforms=, hint them in the query ("on reddit", "from hn and devto"), or restrict by category=. The generated table is on the docs page.

Streaming#

Send Accept: text/event-stream and each platform's result arrives as its search settles, after the plan and before the merged list. Anything refused up front (bad parameters, an unsupported platform, a worst case above your balance) is an ordinary JSON error, never a stream. See Streaming.

What it costs#

• Planning, fusion and the request itself are free.
• Search legs are charged at each platform's normal search price, for the legs that succeeded. A cache hit is free and a failed leg is refunded.
• A worst-case check runs first. If the legs' worst case is above your balance the call is refused before anything runs.
• An Idempotency-Key replays the stored answer for 24 hours at 0 credits.
• One call counts as one request against your rate limit.

1,000 free credits cover the open-platform endpoints. Hard-to-reach platforms (X, TikTok, Instagram, Reddit, and others marked on their page) need a verified email and a purchased pack.

Limits, honestly#

• The planner is string rules, not understanding. A hint needs a preposition ("on reddit") or a content noun ("github repos about x"); a bare "reddit rust" is not a hint.
• Tag-style searches get one keyword. plan.adapted and plan.skipped show what each platform received; a platform no form fits gets no leg.
• Reranking, when enabled, is per platform against the cleaned query. There is no cross-platform model.
• Only search endpoints take part. include and label are not accepted here.
• The hosted MCP server exposes this as one tool, search_universal, without the dry-run option.