Search is the first thing most people add to an agent, and the pricing pages make it look like a coin flip. Seven servers, seven different units of account: per request, per credit, per page, per run, per million tokens. Comparing them honestly means converting all of that into one number and then admitting that the number is rarely what should decide it.
So here is the short version. Two of these are free and need no API key at all. The paid ones cluster between $5 and $8 per 1,000 requests, which is close enough that cost is not the deciding factor between them. And one of the seven is not a search engine, which is the single most common mistake people make when they wire up this category.
- Free with no API key: Fetch (Anthropic official, retrieves a URL) and DuckDuckGo (conventional web search). Start here.
- Cheapest paid: Brave Search and Perplexity, both $5 per 1,000 requests.
- Best free allowance: Tavily, 1,000 credits a month, then $0.008 a credit.
- Best for research: Exa. $7 per 1,000 searches, $20 of credits to start, and an Agent mode that runs whole tasks.
- Firecrawl is not a search server. It bills per page scraped, answers “what is on this page”, and pairs with a search server rather than replacing one.
Prices on this page were read from each vendor’s own pricing page on 24 August 2026. This category changes its pricing more often than most, so treat the figures as a snapshot and check before you commit to volume. ContextBolt does not sell a web search server, so nothing on this list is ours.
Pick one by the job
- You just want the agent to read a page you name. Fetch. Free, official, no key.
- You want ordinary web search and nothing to sign up for. DuckDuckGo.
- You want cheap search at volume. Brave Search, $5 per 1,000.
- You want an agent to actually research something. Exa, or Tavily if you would rather stay on a free tier while you learn.
- You want a cited answer rather than links. Perplexity.
- You need the full content of specific pages, reliably. Firecrawl, alongside one of the above.
How we compared them
Every price below is normalised to cost per 1,000 requests so the numbers are actually comparable. Where a vendor bills in credits, one search is treated as one credit unless their docs say otherwise. Free tiers are quoted as the vendor states them.
What is deliberately not in the ranking: raw index size, which nobody can verify from outside, and benchmark result quality, which depends far too much on your queries to generalise honestly.
Web search MCP servers compared
| Server | API key | Free tier | Paid, per 1,000 | What it returns |
|---|---|---|---|---|
| Fetch (official) | None | Unlimited | Free | One URL as Markdown |
| DuckDuckGo | None | Rate limited | Free | Conventional results |
| Brave Search | Required | $5 credits/mo | $5 | Independent index |
| Perplexity | Required | None | $5 | Grounded answers |
| Exa | Required | $20, then $10/mo | $7 | Cleaned content |
| Tavily | Required | 1,000 credits/mo | ~$8 | Agent-shaped results |
| Firecrawl | Required | 1,000 credits/mo | Per page, not per query | Full page content |
1. Fetch: free, official, and enough more often than you think
Anthropic’s own server. You give it a URL, it returns the page converted to Markdown. There is no search, no key, no account and no bill.
It is on this list because a large share of what people install a paid search server for is actually this job. “Read this documentation page and tell me what changed” does not need a neural index. It needs the page. Installing Fetch first and finding out how far it gets you is the cheapest decision in this whole category.
Where it stops: it cannot find anything. You have to know the URL. It also handles JavaScript-heavy pages poorly, which is precisely where Firecrawl earns its money.
Verdict: install it regardless of what else you pick. Fetch in the directory.
2. DuckDuckGo: search with nothing to sign up for
Conventional web search, no API key, no account. Results are titles, URLs and snippets, the same shape a person would get.
The pairing that makes this work is DuckDuckGo plus Fetch: one finds the pages, the other reads them. Between them they cover most everyday agent research at a cost of zero, and they are the honest recommendation for anyone experimenting rather than building.
Where it stops: there is no paid tier, so when you hit the rate limit there is no lever to pull. Results are also conventional web results rather than the cleaned, ranked output that agent-native APIs return, so deep research work will get further elsewhere.
Verdict: the right first search server for almost everyone. DuckDuckGo in the directory.
3. Brave Search: the cheapest paid option, on its own index
$5 per 1,000 requests, with $5 of credits applied free every month. Brave runs its own index rather than reselling someone else’s, which matters if you care about not routing every agent query through the same two companies.
The free monthly credits are a genuinely nice model: 1,000 requests a month, permanently, without a trial clock. For a personal agent that searches occasionally, the running cost is plausibly zero forever.
Where it stops: results are conventional search results, so you will often want Fetch or Firecrawl behind it to actually read what it found. The Answers tier costs $4 per 1,000 plus token charges if you want grounded responses instead.
Verdict: best value paid search, and the one to pick if index independence matters to you. Brave Search in the directory.
4. Perplexity: answers with citations, not a list of links
$5 per 1,000 search requests. What you get back is a grounded answer with citations rather than results to sift, which is a different product from everything above it on this list.
That is worth paying for when the agent’s next step is reasoning rather than reading. It is worth avoiding when you want the raw sources, because an answer is a summary and summaries lose the detail you were going to check.
Where it stops: there is no free tier, so evaluating it costs money from the first request. The Agent API adds per-tool charges on top, at $0.0025 an invocation for web search plus model token costs, so an agentic workflow bills on more than one meter at once.
Verdict: good when you want conclusions. Reach past it when you want sources. Perplexity in the directory.
5. Exa: the strongest research option, and the most billing surface
$7 per 1,000 searches, with $20 of credits on signup, around 2,800 searches, and $10 of credits added every month after that. Exa is built for agents rather than people: it returns cleaned page content instead of snippets, and its neural index handles the descriptive queries that keyword engines fumble.
The part worth knowing about is Exa Agent, which takes a goal rather than a query and runs the whole research loop, chaining searches and reading pages. It is off by default and it is the single biggest reason to pick Exa. It also bills separately, from $0.012 to $1.00 per run depending on effort level.
Where it stops: the billing surface is wide. Search, contents, answers, monitors and agent runs are all metered separately, and an agent that reaches for Agent mode on a question one search would have answered costs eighty times more. Know which tool it is calling. We cover the whole setup in the Exa MCP guide.
Verdict: the best research server here, if you watch what it spends. Exa in the directory.
6. Tavily: the friendliest free tier
1,000 API credits a month free, then $0.008 a credit, which is about $8 per 1,000. Tavily is built explicitly for LLM workflows and returns results already shaped for an agent to consume.
Its real advantage over Exa is not technical, it is the on-ramp. A thousand credits a month, permanently, is enough to run a genuine research habit without a card on file, and it means you can find out whether agent-native search actually changes your workflow before deciding anything.
Where it stops: slightly more expensive per request than Exa at volume, and without an equivalent of Exa Agent it stays a search tool rather than a research one.
Verdict: the best place to start with agent-native search. Tavily in the directory.
7. Firecrawl: not a search server, and that is the point
This is where the category confusion lives. Firecrawl does not answer “what is out there”. It answers “what is on these pages”, and it bills per page scraped, not per query: 1,000 credits a month free, then $16 a month for 5,000 pages, $83 for 100,000.
That different unit is why comparing it on cost per 1,000 searches is meaningless. It is doing a different job, and it does that job better than anything above: JavaScript-rendered pages, structured extraction, crawling a whole site rather than fetching one URL.
Where it stops: it needs to be told where to look. Discovery is somebody else’s job.
Verdict: not an alternative to a search server, a companion to one. Firecrawl in the directory.
What most people should actually run
The instinct is to pick one. The better answer is usually three, and because everything here bills per request, the ones you barely touch cost you almost nothing.
The free stack: DuckDuckGo for discovery, Fetch for reading. Costs nothing, needs no account, and covers a genuinely large share of everyday agent work. Run it for a fortnight before paying for anything and you will know exactly which limit you actually hit.
The research stack: Tavily or Exa for search, Firecrawl reserved for pages that need proper extraction. Both free tiers together give you 2,000 requests a month at no cost.
The volume stack: Brave Search at $5 per 1,000 for the bulk of queries, Exa kept for the searches where the neural index earns its premium.
The other kind of search, and why it is a separate category
Everything on this page searches the open web. There is a second category that people conflate with it and should not: servers that return data about search rather than results from it. Keyword volumes, ranking positions, SERP composition, backlink profiles.
Exa will tell you what the web says about a topic. It will not tell you where you rank for it, what a term is worth, or who is above you. That is a different data source with different economics, and we compare that category separately in Best SEO MCP Servers, including what ContextBolt SEO does, which is the one we build.
There is a third gap worth naming while we are drawing lines. Web search covers what is public. It does not cover the things you personally decided were worth keeping: the thread you bookmarked, the post you saved, the article you flagged six months ago. No amount of search budget reaches those, because they are not a search problem. That is what ContextBolt is for, and it is the half of an agent’s context that no server on this page can supply.
The honest take
The pricing spread on this page is about $3 per 1,000 requests between the cheapest paid option and the most expensive. For anyone running a personal agent, that is a rounding error, and picking on price is optimising the wrong variable.
What actually differs is shape. Brave gives you links. Perplexity gives you conclusions. Exa and Tavily give you content an agent can use without a second call. Firecrawl gives you a page, properly, when the others cannot render it. Choosing well means knowing which of those four things your agent is short of, and that is a question about your workflow rather than about the vendors.
The one recommendation that holds for everyone: start with the two free ones. Fetch and DuckDuckGo cost nothing, take a minute to install, and the limit you eventually hit will tell you exactly which paid server you need. That is a better signal than any comparison table, including this one.
For the full list of servers in this category and every other, see the MCP Server Directory, and if the protocol itself is new to you, start with what MCP actually is.