Linkup is our first pick for agents whose answers need to be grounded in sources. We like its combination of sourced output, published evaluations, and straightforward request pricing. Tavily is a useful starting point for general retrieval, while TinyFish, Firecrawl, and Context.dev deserve a look when discovery and content extraction need to work together.
There is more than one job here. Finding a page, extracting its contents, and interacting with a website call for different capabilities. This is our shortlist by use case, based on the products’ published offerings. Prices were checked in September 2026.
Our picks at a glance
| Rank | Tool | Where we’d use it | Pricing model |
|---|---|---|---|
| 1 | Linkup | Source-grounded answers and structured retrieval | Standard search $0.005–$0.006 per call |
| 2 | TinyFish | Search followed by live website workflows | Search and Fetch free within limits; Agent and Browser metered |
| 3 | Kernel | Browser sessions for interactive workflows | Free and paid plans, plus compute usage |
| 4 | Context.dev | Search plus page, brand, and structured data extraction | 500 free credits/month; paid plans from $25/month |
| 5 | Tavily | General agent retrieval | 1,000 free credits/month; pay as you go at $0.008/credit |
| 6 | Perplexity Search API | Search alongside Perplexity’s other APIs | Usage-based |
| 7 | Brave Search API | Search from an independent index | Search $5/1,000 requests |
| 8 | Firecrawl | Search, scraping, and site crawling | Credits and subscription plans |
| 9 | Serper | Google result positions and snippets | Prepaid query packs |
Kernel is a complementary browser service in this list. It isn’t a substitute for a search index. Google’s Custom Search JSON API is also worth flagging for existing users: it is closed to new customers, so we would not shortlist it for a new build.
1. Linkup: our pick for source-grounded answers
Linkup gives teams a choice between search results, sourced answers, and structured output. That makes it appealing when a product needs both readable answers and fields it can use downstream. Its published benchmark reports a 94% F1 score for Linkup Fast on its SimpleQA evaluation. We see that as a reason to test it, rather than a guarantee about your application’s accuracy.
The billing documentation lists standard search at $0.005 for search results and $0.006 for sourced or structured output. Deep search costs $0.05 and $0.055 respectively. Errors and queries with no relevant results aren’t charged; an exhausted prepaid balance stops requests with an HTTP 429. The docs also describe a $20 signup balance replenished to $20 monthly.
One distinction matters for procurement: Linkup’s pricing page includes SOC 2 Type II on all plans but lists zero data retention under Enterprise. They are separate requirements, and the plan needs to match yours.
Our take: start here when source quality and the shape of the answer matter more than matching a consumer search results page. Test citation support and failure behavior with your own questions before committing.
2. TinyFish: search that can lead into a website workflow
TinyFish offers Search, Fetch, Agent, and Browser products. That combination is interesting when finding the URL is only the first step: the task might continue through a dynamic page or a multi-step interaction.
Its pricing page lists free Search at up to 30 requests per minute and 500 per hour, and free Fetch at up to 150 URLs per minute and 1,000 per day. Agent steps and browser time are billed separately.
Our take: evaluate it for workflows that span discovery, extraction, and interaction. Measure completion of the whole task, including the paid steps, rather than comparing only the search price.
3. Kernel: browsers for the steps search cannot complete
Kernel supplies browser infrastructure for agents and automation. It is relevant when your workflow needs an authenticated session, a page interaction, or a browser you can inspect and replay.
We would pair it with a search provider when needed. Its pricing combines plan features with usage measured in GB-seconds, so cost depends on browser compute rather than a fixed search-query count. Evaluate session reliability and recovery as well as runtime.
4. Context.dev: search plus structured web context
Context.dev combines web search with scraping, crawling, and extraction of content such as brand information and images. We’d shortlist it for enrichment or research features that need multiple kinds of information from a company website.
Its pricing page lists 500 free monthly credits and a $25/month Developer plan with 10,000 credits. Web Search is listed at one credit per ten results; other operations have their own costs. A brand retrieval and a page scrape should not be counted as equivalent work.
5. Tavily: a practical general-purpose option
Tavily combines search with extraction, crawling, mapping, and research products. We would consider it for a team that wants several retrieval tasks under one provider. Its pricing page confirms 1,000 free API credits per month and a $0.008 pay-as-you-go credit rate.
Credits aren’t interchangeable with searches: the operation and request settings determine consumption. Budget using the mix of calls your agent actually makes, rather than dividing the free allowance by one assumed request cost.
6. Perplexity Search API: search alongside an answer API
Perplexity’s API platform offers both search and grounded answer products. We would look at Search API when our application needs retrieved sources for its own model, and Sonar when we want Perplexity to compose an answer.
Those are different purchasing decisions. Compare the cost of the complete workflow, including any model generation, rather than treating the two products as equivalent requests.
7. Brave Search API: an independent index
Brave operates its own search index and offers search results with additional context for LLMs. Its Search plan lists $5 per 1,000 requests and $5 in monthly free credits.
We like it as a candidate when index independence matters. Its AI context offering also means it shouldn’t be dismissed as a provider of bare links. Compare the returned passages against the context your model needs.
8. Firecrawl: discovery and extraction together
Firecrawl supports web search as well as scraping, mapping, and crawling. We’d consider it when an agent needs to discover pages and then collect substantial content from them, or when the same product also ingests documentation sites.
For budgeting, distinguish a search request from the pages subsequently scraped. A workflow that expands into a whole site has different economics from retrieving a few passages.
9. Serper: when the search results page is the data
Serper provides Google search result data. We’d use it for rank monitoring and applications that need result positions and snippets. It can also feed an agent, but decide whether snippets are sufficient or whether you need to fetch the linked pages.
The relevant distinction is how much source content you receive, not whether an API can return structured data at all.
How we’d choose
Start with a set of real product questions and known-good answers. Keep the model and prompt fixed, then compare final answer correctness, citation support, latency, and cost. Include recent events, ambiguous names, and questions with no reliable answer. A useful system needs to handle those without inventing certainty.
For tools that browse or extract, add a separate task-completion test. Can the workflow reach the right page, collect the required fields, and recover from a failure? A search benchmark alone won’t answer that.
Public web context is also only part of many B2B agent workflows. Account ownership, open deals, and billing state come from customer systems. Our posts on connecting agents to GTM data, adding revenue and billing context, and mapping customer-specific fields cover that side of the architecture.