TL;DR
- Agents search a lot: 5–50 searches per task is normal. Price per 1,000 and latency matter far more than for a chatbot.
- The hidden cost is tokens. Search results your model reads can cost more than the search itself.
- MCP servers plug a search API into Claude Code, Cursor, Codex and other agent tools without writing glue code.
- For agent loops: Keirolabs (lite from $0.03, fast from $0.44, MCP server, JS-rendering extract), Parallel (compressed excerpts, $1 fast) and Exa (page text).
A chatbot searches once and answers. An agent searches, reads, decides it needs more, searches again, opens a page, compares, and searches again. Five to fifty searches per task is normal, and a coding agent researching a library can do more.
That changes what “best search API” means. Here’s what matters for agents, how MCP search servers fit in, and which APIs are built for it.
What an agent needs from a search API
- Low price per call. At 20 searches a task and 50,000 tasks a month, you’re making a million searches. $5 per 1,000 is $5,000; $0.10 is $100.
- Low latency. Every search sits in the agent’s critical path. A 3-second search times 20 is a minute of waiting.
- Few tokens, high relevance. The model pays to read whatever you return.
- Page content on demand. Snippets to decide, full text when it matters.
- High rate limits. Agents run in parallel. A 1-request-per-second cap breaks a fan-out.
- An easy integration. Ideally an MCP server, so any agent tool can use it.
The real cost: tokens, not searches
Say each search returns 3,000 tokens of text and your model charges $3 per million input tokens. Reading the results costs $9 per 1,000 searches, more than most searches. In a multi-turn agent, earlier results stay in context and get re-read every turn, so multiply again. Cutting returned tokens by half saves more than switching to a cheaper search API.
That’s why the best agent search APIs give you control over how much text comes back:
- Snippets first. Use a cheap snippet search to decide which pages are worth reading.
- Compressed excerpts. Parallel returns ranked, compressed excerpts instead of whole pages.
- Evidence windows. Keirolabs’ deep search returns ranked evidence from top pages; its public benchmark found the answer in the ~900-token evidence alone 94.0% of the time.
- Filter before context. Claude’s newer web search tool versions can filter results in code before they reach the context window.
What is a web search MCP server?
MCP (Model Context Protocol) is a standard way to give AI tools access to outside capabilities. A search MCP server exposes a search API as tools the agent can call by itself: search, extract, answer and so on.
Instead of writing code that calls the API and pastes results into a prompt, you add the server to Claude Code, Cursor, Codex or any MCP client, and the agent decides when to search.
Most AI search providers now ship one:
| Provider | MCP server | What the agent gets |
|---|---|---|
| Keirolabs | Remote MCP | Every endpoint: lite/fast search, content, extract, answer, deep, Google SERP, batch |
| Exa | Yes | Search, contents, research |
| Tavily | Yes | Search, extract, map, crawl |
| Brave | Yes | Web, news, images, local |
| Parallel | Yes, plus a free Search MCP | Search, extract |
| Firecrawl | Yes | Scrape, crawl, search |
With an MCP server, cost control moves into your prompt. Tell the agent to use the cheap search first and only open pages it needs. Otherwise it will happily fetch full content for every result.
The best search APIs for agent loops
Prices per 1,000 calls from each vendor’s pricing page, November 5, 2026.
| API | Cheapest search | With page text | Answers | Notes |
|---|---|---|---|---|
| Keirolabs | $0.03–0.10 (lite) | $1.33–4.00 | $0.89–2.67 | Own index, MCP, batch, JS-rendering extract |
| Parallel | $1 (turbo/fast) | Excerpts included | $10+ (Responses) | Token-efficient excerpts |
| Exa | $4 (lite) | + $1 per 1,000 pages | $5 | Own neural index |
| Perplexity Search | $1 (fast) | No | Sonar/Agent API | Own index |
| Tavily | $5–8 | Included | Included | Aggregated |
| Brave | $5 | LLM context endpoint | $4 + tokens | Independent index |
| OpenAI / Claude built-in | $10 + tokens | Included | It’s the model | Locked to that model |
Why Keirolabs is the one I’d start an agent on
Agents are where cheap search matters most, and nothing we track is cheaper. At $0.03 per 1,000 on Startup, a million lite searches uses 37,500 credits, a third of the $99.99 plan. The rest of an agent’s toolkit sits on the same key and credit balance:
- Fast search (0.5 credits) for richer results, search + markdown (1.5) for RAG
- Extract (1 credit) to read one URL, answer (1 credit) for cited answers
- Deep search (2 credits): 96.5% on SimpleQA, code and results public
- Batch search, free and unlimited on Startup, for fan-out
- Extract with JavaScript rendering (deep and heavy modes) when the agent hits a page search can’t read
- Rate limits up to 600 lite searches a minute on Startup, and retries within a request aren’t charged twice
The pattern I’d ship: lite search to triage, extract or content only on the 2–3 pages that matter, deep search for the hard questions. On Keirolabs that whole loop costs a fraction of one Tavily search. 750 free credits to try it, and KEIRO20 takes 20% off through Black Friday: pricing.
Parallel: the token saver
If token bills dominate, Parallel’s compressed excerpts are the best-designed answer. Fast is $1 per 1,000.
Exa: neural search with page text
Exa’s embedding-based index is excellent for “find things like this” queries, and contents cost $1 per 1,000 pages. See Exa API pricing.
Built-in model search: easiest, priciest
OpenAI and Claude’s built-in web search is one line to turn on, at $10 per 1,000 plus tokens. Good for prototypes, expensive in a loop.
A checklist before you pick
- Log a day of real agent traces and count searches per task.
- Measure tokens added per search with each candidate.
- Test p95 latency with parallel calls, not one at a time.
- Check the rate limit on the plan you’d actually buy.
- Price it: (searches × search price) + (tokens × model price). The calculator does the first half.
FAQ
What’s the cheapest web search API for agents?
Keirolabs lite search: $0.10 per 1,000 on Essential, $0.06 on Pro, $0.03 on Startup.
Do I need an MCP server?
No, but it saves time. Any API can be wrapped as a tool in your own code. MCP is the shortcut for tools like Claude Code and Cursor.
How many searches does an agent make?
It varies widely. Simple Q&A agents make 1–5; research and coding agents often make 20–50. Measure yours.