# Which web search API has a free tier for building an AI agent? > Source: > Published: 2026-11-15 15:08:46+00:00 Linkup has the largest monthly free tier of the web search APIs in this comparison: $20 of credit a month, about 4,000 standard searches, for an account with a work email. Parallel gives up to 5,000 searches a month, but only in its `fast` mode and with a card on file. Of the monthly tiers that need no card and accept any email, Exa's $10 a month, about 1,400 searches, is the largest. To start with no account at all, Exa, Tavily, Firecrawl and Parallel run keyless MCP servers, with limits they do not publish. Each row comes from the vendor's own pricing page or docs, read on November 14, 2026. "Searches" is the count in the default search mode with 10 results, unless the row says otherwise. | API | Free allowance | Refill | Searches it buys | Card needed | Rate limit stated | |---|---|---|---|---|---| | [Linkup](https://docs.linkup.so/pages/documentation/development/pricing) | $20 of credit, for an account with a work email | Topped up to $20 each month | About 4,000 at $0.005 a standard search | Not stated | Not stated | | [Parallel](https://parallel.ai/pricing) | $5 of credit, plus up to $80 at sign-up | $5 monthly; the $80 one time | 1,000 in the default `advanced` mode, 5,000 in`fast` | Yes, a card on file for the $5 | 600 a minute | | [Exa](https://exa.ai/pricing) | $10 of credit, plus a one-time $10 for onboarding | Resets to $10 on the 1st | About 1,428 `auto` , 2,500`lite` | No | 10 a second | | [Brave](https://brave.com/search/api/) | $5 of credit on the Search plan | Monthly | 1,000 | Yes, "to prevent abuse" | 50 a second | | [Tavily](https://www.tavily.com/pricing) | 1,000 credits | Monthly | 1,000 basic, 500 advanced | No | 100 a minute | | [Firecrawl](https://www.firecrawl.dev/pricing) | 1,000 credits | Monthly | 500 (2 credits per 10 results) | No | 10 searches a minute, 2 at a time | | [Serper](https://serper.dev/) | 2,500 queries | At sign-up; not stated as monthly | 2,500 | No | Not stated | | [SerpApi](https://serpapi.com/pricing) | 250 searches | Monthly | 250 | Not stated | 50 an hour | | cloro | 500 credits | Monthly | 166 Google searches, or 100 with the AI Overview | No | 1 request at a time | | [Perplexity Search API](https://docs.perplexity.ai/getting-started/pricing) | None stated | | 0 | | 50 a second | Three rows need a closer read. Brave's banner says "Up to $20 in free calls every month", but its Search plan carries $5 a month, which buys 1,000 searches. Parallel's FAQ says "every account gets a recurring free monthly allowance of $5", and its plan card says the $5 needs a card on file. Linkup's top-up applies to "eligible accounts"; its docs name the work email and no other condition. The cloro row in numbers: a Google search is 3 credits and the AI Overview makes it 5, so 500 credits buy 166 or 100 searches on the async API. A synchronous request adds 2 credits, which makes it 100 or 71. Four vendors run a hosted MCP server that answers without an API key: `https://mcp.exa.ai/mcp`. Search and fetch "at a free rate limit". It needs a direct IPv4 connection; from IPv6 or a proxy, you need a key or OAuth.`X-Tavily-Access-Mode: keyless` to `https://api.tavily.com/search`, `/extract`, or the MCP server at `https://mcp.tavily.com/mcp/`. Crawl, map and research need a key. None of the four publishes a number for its keyless limit. Keyless access is enough to wire up a prototype in Claude Code or Cursor; move to a key before the agent searches in a loop. Open-source servers such as `duckduckgo-mcp-server` scrape a search engine from your own machine with no key, and break when the engine changes its page. Each free tier returns the same response as the vendor's paid plans. The source and the shape differ by vendor. | API | Source | One search returns | |---|---|---| | Linkup | Linkup's index | Results, or a sourced answer for $0.006 | | Parallel | Parallel's index | Results with excerpts from each page | | Exa | Exa's neural index | Results with page text or highlights | | Brave | Brave's own index | Results with snippets, as JSON | | Tavily | Tavily's index | Results with extracted page content | | Firecrawl | Its own search endpoint | Results; page content costs more credits per result | | Serper | | The Google results page as JSON | | SerpApi | Google and other engines | The Google results page as JSON | | cloro | | The results page as JSON, with the AI Overview and People Also Ask | | Perplexity Search API | Perplexity's index | Ranked results with snippets | Page text (Exa, Tavily, Linkup, Parallel) is what a model answers from, and it costs tokens on every call. Structured results (Brave, Serper, SerpApi) let the agent read a field, such as a rank or a URL. An agent seldom stops at one search. Assume a research agent that runs 5 searches per task. The table divides each monthly allowance by 5, then by 30 days. | API | Searches a month | Tasks a month | Tasks a day | |---|---|---|---| | Parallel, `fast` mode | 5,000 | 1,000 | 33 | | Linkup | 4,000 | 800 | 26 | | Exa, `auto` | 1,428 | 285 | 9 | | Brave | 1,000 | 200 | 6 | | Tavily, basic | 1,000 | 200 | 6 | | Parallel, `advanced` | 1,000 | 200 | 6 | | Firecrawl | 500 | 100 | 3 | | SerpApi | 250 | 50 | 1 | | cloro, async | 166 | 33 | 1 | Serper's 2,500 queries are a one-time grant: 500 tasks in total. Two limits change these numbers in practice. **Concurrency.** An agent that fires 5 searches at once waits on a tier that runs 1 request at a time, and SerpApi's 50 searches an hour cap it at 10 tasks an hour. Exa, Brave, Tavily and Parallel leave room for parallel searches. **Misses.** Six web search APIs ran on the 824 multi-hop questions of the FRAMES set in November 2026, with one search of 10 results per question. The two Google SERP APIs answered 80.2% and 77.4% of them. Parallel answered 65.3%, Exa 64.0%, Brave 53.9% and Tavily 48.4%. Those four replied "not found" on 151 to 295 questions, and an agent would search again on each of them. On short factual questions the order changes: Parallel (97.8%), Exa (97.3%) and the top Google API (96.9%) are within the margin of error. The questions, settings and graded answers are in the [benchmark repo](https://github.com/cloro-dev/web-search-api-benchmark). | Your agent needs | Free tier to start on | |---|---| | The most searches a month | Linkup, or Parallel in `fast` mode with a card | | No card and any email | Exa, then Tavily or Firecrawl | | No sign-up, for a prototype | A keyless MCP server: Exa, Tavily, Firecrawl or Parallel | | An index that is not Google | Brave, Exa or Parallel | | Google's own results page | Serper for volume; SerpApi or a Google SERP API with the AI Overview for the full page | A free tier tests the response shape and the tool wiring. Past a few thousand searches a month, the paid rate and the share of searches that hold the answer decide the cost. This comparison is published by cloro, which is one of the APIs in it. Every other allowance, refill rule, card rule and rate limit comes from that vendor's own pricing page or docs, read on November 14, 2026, and each one is linked in the first table. The cloro row comes from its published plans and billing docs. Serper's page does not say whether its 2,500 queries refill; third-party reviews describe them as one-time. The accuracy figures are copied from the benchmark linked above. The full comparison, with paid prices and latency for each API, is in [Best Web Search API for AI Agents](https://cloro.dev/blog/best-search-api/).