Author(s): Jamieparker

Originally published on Towards AI.

I spend most of my time thinking about how AI agents connect to the outside world. I watch what breaks, what costs more than expected, and what vendor changes force teams to rewire their setups at the worst possible time.

Search is the most common point of failure. Agents that need live data, current prices, recent events, or anything past the model’s training cutoff depend entirely on a search API being where you left it. And in 2026, that has been less reliable than most teams expected.

This piece covers seven tools I have direct experience with or have tracked closely, what each one actually returns, what it costs today, and what happened to it in the past six months that the vendor’s marketing page will not tell you.

Why your agent’s search tool matters more than its model

An agent that quotes a price, recommends a product, or names who runs a company right now is not reasoning about any of that. It is repeating whatever was true when its training data ended, with no built-in way to flag that anything has changed.

A search tool closes that gap mid-task. But not all of them do the same job.

The tools in this list split into three distinct capabilities:

Finding pages: Returns ranked URLs or snippets. The agent decides what to do with them.

Reading content: Returns full page text alongside search results, skipping a separate scraping step.

Producing a finished answer: Synthesizes results into a sourced response in a single call.

Most tools do one or two. One does all three. Knowing which capability you need is the fastest way to cut this list to one or two candidates.

The seven tools, what I found

1. Exa

Exa is a neural search API built specifically for AI workloads. It ranks pages by meaning rather than exact keyword match, and as of a March 2026 update, returns page content by default in the same call.

In May 2026, Exa raised $250 million in a Series C at a $2.2 billion valuation, led by Andreessen Horowitz, with Benchmark and Y Combinator participating. That matters not because of the number but because of what the hiring spree it funded tells you about where the product is going: the former head of retrieval infrastructure at Meta, the head of search backend at Yandex, and a research team from Google. They are building a different kind of search infrastructure than what Google designed for humans clicking links.

What I like about it: Neural ranking finds relevant pages even when the source never used your query’s exact wording. That is the difference between a research agent that surface-skims and one that actually finds the right paper. The Exa Agent tool runs multi-step research through the same MCP server connection used for regular search, so there is no second product to wire in.

What to watch: No volume discount without an enterprise conversation. The default rate limit sits at 10 requests per second, which parallel agent pipelines hit faster than expected. I have seen teams get caught by this mid-demo.

Pricing: Free (20,000 requests/month), $7 per 1,000 search requests, $1 per 1,000 pages for content, $12–15 per 1,000 for Deep Search.

Best fit: Research-heavy agents working with conceptual queries where keyword matching would miss what they need.

2. Tavily

Tavily is the search API I see most often inside LangChain and LlamaIndex-based RAG pipelines. It returns cleaned, model-ready output rather than raw HTML, with a credit system that makes budgeting straightforward: basic search costs 1 credit, advanced costs 2.

The acquisition is the biggest thing to know right now. In February 2026, Nebius Group acquired Tavily for $275 million, with total consideration potentially reaching $400 million on performance milestones. Tavily’s founder joined Nebius and the product continues under its own brand. For teams with long-term dependencies on the API, the roadmap and enterprise pricing questions are worth confirming directly rather than assuming.

What I like about it: Search, extract, map, and crawl live under one API. The include_answer parameter returns a short synthesized answer alongside raw results, skipping a second endpoint for simple lookups. Remote MCP via OAuth means no local install. Tight native integrations with LangChain and LlamaIndex mean setup is a few lines of code rather than a custom wrapper.

What to watch: Heavy structured extraction beyond snippets usually still needs a dedicated scraping tool alongside it.

Pricing: Free researcher tier, $0.008 per credit pay-as-you-go, $30/month Project tier (roughly 4,000 basic searches), enterprise custom.

Best fit: LangChain and LlamaIndex pipelines that need predictable per-search pricing and clean model-ready output without building a custom retrieval layer.

3. MCP360

MCP360 is a gateway that connects an agent to 100+ external tools, including several of the APIs in this list, through a single configuration entry. Search, SERP tracking, keyword research, web scraping are all wired through one credential and one response schema instead of five.

The reason I keep coming back to this is what happened in February 2026 when Brave restructured its free tier. Teams using Brave through MCP360 did not touch their agent setup. The routing change was handled at the gateway layer. Teams using Brave directly had to update credentials, retest their tool-call logic, and in some cases rebuild the pipeline when the response schema changed.

That is the tradeoff in one example. One vendor change at the gateway level versus one vendor change per team per integration.

What I like about it: One response schema, automatic failover when an upstream API changes, on-demand tool loading so a 100-tool catalog does not inflate the agent’s context window. The Custom MCP Builder wraps any REST API into an MCP tool without code for standard APIs.

What to watch: Overhead for teams that genuinely need only one search tool. The free tier is for testing, not production.

Pricing: Free (100 credits/month), Starter $16/month, Professional $83/month, Advanced $333/month.

Best fit: Agent stacks with five or more external tool dependencies where managing each vendor’s credentials, schemas, and failure modes individually starts eating real engineering time.

4. Firecrawl

Firecrawl pairs a search endpoint with full-page extraction in one workflow. Find a page, read it, all under one provider and one credit pool. No second vendor for scraping.

The token efficiency argument is concrete: Firecrawl’s own benchmark puts clean markdown at roughly 2,788 tokens compared to 38,381 tokens for the same raw HTML page. In a high-volume agent pipeline, that difference is not marginal.

What I like about it: JavaScript rendering is on by default, so pages requiring a real browser still return usable content. The MCP server ships five core tools (scrape, crawl, map, search, extract), an autonomous research agent, and page-change monitors. Core tools work without an API key on the free tier, rate-limited but instantly usable for testing.

What to watch: Credits do not roll over on the standard plan except for auto-recharge credits and annual Scale tiers. JSON extraction and Stealth Mode consume credits several times faster than a plain scrape. A pipeline using both regularly can burn through credits faster than the plan sizing suggests.

Pricing: Free, Hobby $16/month, Standard $83/month, Growth $333/month, Scale $599/month, Enterprise custom.

Best fit: Agents that need to search for pages and then read their full content in one continuous workflow. If your agent finds results and then has to read them, this cuts the complexity and the vendor count.

5. Perplexity

Perplexity gives developers two distinct products under one account: a raw-results Search API and the Sonar model family, which combines web retrieval and LLM synthesis with citations in a single call.

The thing that catches teams off guard: the developer API is completely separate from the consumer Perplexity Pro subscription. Paying for Pro does not give you meaningful API access. I have seen at least three teams discover this after they assumed the Pro account covered their agent integration.

What I like about it: Sonar removes the find-read-synthesize step entirely. For agents that need sourced answers without building a RAG loop, this is the fastest path. The endpoint is OpenAI-compatible, so switching the base URL is the entire integration. Four purpose-built MCP tools (perplexity_search, perplexity_ask, perplexity_research, perplexity_reason) let an agent route easy lookups to the cheaper tier and deeper research to Sonar Pro.

What to watch: Every Sonar call carries a per-request fee on top of token cost. Teams accustomed to token-only billing find the actual cost higher than the listed token rates suggest. Sonar also controls which sources get cited, so agents that need auditable, step-by-step retrieval are better served by the standalone Search API paired with a separate model.

Pricing: Free tier, Search API $5 per 1,000 requests, Sonar $1/M input + $1/M output + request fee, Sonar Pro $3/M input + $15/M output + request fee.

Best fit: Agents that need cited, synthesized answers in one call and do not require control over the retrieval path.

6. Brave Search API

Brave runs its own independent web index rather than reselling Google or Bing data. After Microsoft retired the Bing Search API on August 11, 2025, Brave became the only major independent Western search index available to developers at scale. For any team that needed a non-Google, non-Bing source, Brave was suddenly the default.

Two things changed in early 2026 that are worth knowing before you sign up.

First, Brave eliminated its free tier in February 2026 and replaced it with $5 in monthly credits, enough for roughly 1,000 queries. Second, keeping that credit now requires attributing Brave Search API in the project’s public website or about pages. The old free plan had no attribution requirement. The credit card Brave collected at signup is now an active billing instrument, not just identity verification.

What I like about it: The LLM Context endpoint returns structured content chunks and JSON-LD schemas built for a model to read directly. The Answers endpoint returns a grounded, synthesized response through an OpenAI-compatible interface. Query data protections are documented at every tier.

What to watch: The core plan returns search results without full-page content extraction. Reading a complete page still requires a second tool. Attribution is required on all metered plans, which occasionally creates friction for internal tooling where “about pages” do not exist in the usual sense.

Pricing: Search $5 per 1,000 requests, Answers $4 per 1,000 queries plus token costs, Enterprise custom. Attribution required.

Best fit: Privacy-sensitive workloads, regulated industries, and teams that need an index genuinely independent of both Google and Bing.

7. SerpAPI

SerpAPI scrapes structured results from Google, Bing, YouTube, Baidu, and roughly a dozen other engines and returns them as JSON. No synthesis, no content cleaning. The agent’s pipeline handles parsing.

The legal situation is the most important thing to disclose here. Google sued SerpAPI on December 19, 2025, alleging that SerpAPI circumvented Google’s SearchGuard anti-scraping system on billions of separate occasions. SerpAPI filed a motion to dismiss in February 2026, arguing Google lacks standing to bring DMCA anti-circumvention claims over publicly visible search results. A hearing was scheduled for May 19, 2026 before Judge Yvonne Gonzalez Rogers in the Northern District of California. No ruling has been publicly confirmed as of this writing.

I am not saying avoid SerpAPI. The motion to dismiss arguments are substantive, and courts have been skeptical of DMCA anti-circumvention claims tied to publicly visible data in recent cases. But if you are building an agent that depends on it long-term, track the docket. An adverse ruling or injunction mid-pipeline is not a theoretical risk.

What I like about it: Multi-engine coverage across a dozen-plus engines is something no other tool in this list comes close to. Only successful searches draw down the monthly quota. Cached results and failed requests do not waste paid capacity. The MCP server code is public on GitHub, so you can inspect or self-host it.

What to watch: Unused searches do not roll over between billing cycles. The output is raw SERP JSON and requires a parsing step that the AI-native tools here skip. Reddit also filed a separate lawsuit against SerpAPI in October 2025, which adds a second active legal front.

Pricing: Free (250 searches/month), Starter $25/month, Developer $75/month, Production $150/month (~15,000 searches), Big Data $275/month, Enterprise custom.

Best fit: SEO monitoring agents and rank tracking across multiple engines at country level. Not the right fit for agents that need model-ready output.

How I would actually choose

Before applying any of these criteria, I would verify three things directly on each vendor’s current pricing page rather than from a blog post, including this one. Two of the seven tools above changed their free-tier terms in the past six months. Pricing screenshots are perishable.

Semantic or conceptual research: Exa. Neural ranking is meaningfully different from keyword search for this class of query, not just in positioning but in actual result quality.

Search plus full-page reading in one call: Firecrawl. Covers both under one credit pool and cuts one vendor dependency.

Cited answers without building retrieval from scratch: Perplexity Sonar. One call, sourced response. The request fee is real and adds up, so model the actual cost before committing to scale.

An independent index outside Google and Bing: Brave. The only option here that owns its own index. Read the attribution terms before deploying.

Multi-engine SERP data for rank tracking: SerpAPI, with awareness of the active Google lawsuit.

Flat per-search pricing for predictable RAG budgets: Tavily. The credit model is the simplest to budget against at scale.

Five or more external tools behind one agent: MCP360. That is where a gateway layer earns its overhead. Managing five separate credentials, schemas, and failure modes individually is manageable at two tools and genuinely expensive at five.

My Take

Honestly, the “best” search tool is entirely determined by what your agent needs to return and the scale you’re operating at.

That said, if I had to steer someone who’s just getting started, don’t try to wire in three tools at once. Pick the one that maps to your agent’s primary output, get it working, and only add more when you hit a real gap.

Where I do have a stronger opinion is if you’re already running multiple tools like search, scraping, SERP tracking, keyword research, etc, managing each one separately is the kind of low-visibility overhead that quietly drains engineering time. That’s where something like MCP360 starts to make genuine sense, not just as a convenience, but as the simpler architecture. One connection point, one credential to manage, and if an upstream vendor changes something, you’re not the one firefighting it.

It’s not the first thing I’d reach for on day one. But it’s worth keeping in mind as your stack grows.

Join thousands of data leaders on the AI newsletter. Join over 80,000 subscribers and keep up to date with the latest developments in AI. From research to projects and ideas. If you are building an AI startup, an AI-related product, or a service, we invite you to consider becoming a sponsor.

Published via Towards AI