# Customer Spotlight: Seltz, a Fast Web Search API for LLMs and AI Agents
Seltz is a web search API built for LLMs and AI agents. Instead of returning a page of links for a person to click, it returns ranked documents with their full text, shaped to drop straight into a model's context, from an index Seltz builds and owns. In Seltz's own company search benchmark it returned results in 327ms at the median, and the Search endpoint costs $3 to $5 per 1,000 searches depending on the tier. This spotlight covers what Seltz offers, how to connect it to an agent in a few minutes, what its published benchmark does and doesn't show, and how it fits alongside a live search results page from a specific location.
> **Key Takeaways**
>
> - Seltz serves "web search, cited answers, multi-step research, page fetching, and recurring monitors" from "Seltz's own index rather than a live scrape" ([Seltz Docs](https://docs.seltz.ai/llms.txt)).
> - Five endpoints cover different jobs: **Search** (raw documents), **Answer** (one cited response), **Agent** (multi-round research), **Fetch** (content for URLs you already have), and **Monitor** (what's new on a schedule).
> - Search has two tiers: `pro` (default, reranked, $5 per 1,000) and `base` (faster, $3 per 1,000). New accounts get $20 of free requests ([Seltz pricing](https://docs.seltz.ai/pricing)).
> - In Seltz's own run of Exa's company search benchmark, Seltz posted a 327ms median latency and 0.53 precision at 10, while Exa led on top-1 accuracy, 0.68 to 0.63 ([Seltz](https://seltz.ai/blog/company-search-benchmark)).
> - Use an indexed API like Seltz for fast, citable web knowledge. Use a live results page from a set location when the question is "what does a searcher in Chicago see right now."
## What is a web search API for LLMs?
A web search API for LLMs takes a query and returns results a model can read directly: clean text, a URL to cite, and often a date, instead of HTML built for people. It is the retrieval step in retrieval-augmented generation (RAG), as our guide to [building a RAG pipeline on live web data](https://www.joinmassive.com/blog/building-a-rag-pipeline-on-live-web-data-without-stale-indexes) explains. The difference matters in an agent loop. Every extra step to fetch, render, and clean a page adds latency and tokens, and an agent may run dozens of searches to finish one task.
That loop is the design center. Each Search result comes back as a `url` and `content`, with a `published_date` on time-stamped scopes such as news, "rather than raw HTML, so you can pass results directly into a prompt or pipeline without additional cleanup" ([Seltz Docs](https://docs.seltz.ai/search/search-concepts)). Content is capped at 20,000 characters per document by default, and you can lower the cap, switch content off, or ask for short snippets instead.
## Who is Seltz?
On June 24, 2026, Seltz announced a $12.5 million seed round led by Speedinvest and B Capital, with Italian Founders Fund, United Ventures, and Bain & Company's venture arm participating ([Dealroom](https://dealroom.co/news/135823-seltz-raises-12-5m-seed-to-rebuild-web-search-for-ai-agents/)). Seltz is a Massive customer. Everything below about Seltz's product, prices, and performance comes from Seltz's own published docs and blog, and we attribute their numbers to them.
## What can you do with Seltz?
The API is organized by job, and the docs tell an agent which endpoint to call: "Use Answer for one synthesized, cited response; Use Search for the raw documents to reason over yourself; Use Agent for in-depth research needing several rounds of searching" ([Seltz Docs](https://docs.seltz.ai/llms.txt)).
| Endpoint | What it returns | Price (base / pro) | Typical latency |
|---|---|---|---|
| Search | Ranked documents with full text | $3 / $5 per 1,000 searches | About 200ms (base) to 500ms (pro) |
| Answer | One synthesized answer with citations | $6 / $7 per 1,000 | About 7 to 14s |
| Fetch | Clean markdown for up to 20 URLs per call | $5 / $10 per 1,000 URLs | 1 to 25s |
| Monitor | Only what's new on a saved search, on a schedule | $3 / $5 per 1,000 checks | Scheduled |
| Agent | Multi-round research, Markdown or JSON | $0.25 to $2.00 per run | 1 to 30+ min |
Source: [Seltz pricing](https://docs.seltz.ai/pricing), retrieved October 19, 2026.
Search is priced per search, not per document returned, so pulling 10 results instead of 3 doesn't change the bill. The `base` tier makes sense when you re-rank results yourself or aggregate over many of them. Seltz says `base` gives "a faster response, with less precise ordering near the top," and that both tiers search the same data ([Seltz Docs](https://docs.seltz.ai/search/search-concepts)).
Searches also take a `scope`, which picks a slice of the index. The `web` scope (in beta) is the default. The others are companies, people, news, financial reports, and Wikipedia ([Seltz Docs](https://docs.seltz.ai/data/data-concepts)). A `companies` search returns one company per result, which is why the benchmark below uses company queries.
Two design choices stood out to us when we read the docs. First, an unknown search tier is rejected with a `400` instead of quietly falling back to the default, so "the tier you are served is always the tier you named" ([Seltz Docs](https://docs.seltz.ai/search/guides/choose-a-tier)). Second, in Fetch, "a failed page is not a failed call" ([Seltz Docs](https://docs.seltz.ai/fetch/concepts)): a page that can't be fetched comes back inside a `200` with its own error code, so an origin's 404 stays separate from an API failure. Both make an agent's error handling simpler.
## How do you connect Seltz to an AI agent?
The fastest path is the hosted MCP (Model Context Protocol) server at `https://mcp.seltz.ai/mcp`. In Claude Code it is one command, then `/mcp` to authenticate with your API key ([Seltz MCP docs](https://docs.seltz.ai/integrations/mcp)):
BASH
claude mcp add --transport http \
seltz https://mcp.seltz.ai/mcp
According to the [MCP docs](https://docs.seltz.ai/integrations/mcp), it is also listed in Claude's connector directory, with setup steps for Cursor, VS Code, Codex, and Windsurf. For code, the Python and TypeScript SDKs read a `SELTZ_API_KEY` environment variable. A Python search looks like this:
PYTHON
from seltz import Seltz
client = Seltz()
res = client.search(
"best ai search engines",
max_results=10,
)
for doc in res.documents:
print(doc.url)
Tool-calling setups are documented for the Anthropic, OpenAI, and Mistral APIs, plus integrations for LlamaIndex, Agno, Langroid, n8n, Dify, and Zapier.
We installed the official Python SDK (version 1.17.0) on October 19, 2026 to check the snippet against the code itself. The `search` method takes the query plus keyword options for `max_results` (default 10), `scope`, `tier`, `include_domains`, `exclude_domains`, `from_date`, `to_date`, and `fields`, and it returns a response whose `documents` list matches the snippet above.
One detail worth knowing: the SDK doesn't call the REST endpoint. It opens a TLS gRPC channel to `grpc.seltz.ai` with keepalive pings, while the docs' curl examples use REST at `api.seltz.ai`. For an agent making many calls, a long-lived channel avoids a new connection per search. If your network blocks gRPC, the docs also document a REST Search endpoint at `api.seltz.ai`.
## How fast and accurate is Seltz?
Seltz's headline figure is a 327ms median on company search, from a September 2026 benchmark post ([Seltz](https://seltz.ai/blog/company-search-benchmark)). Its [homepage](https://www.seltz.ai/) also says it retrieves full web pages in under 200ms.
That benchmark is worth reading closely. Seltz ran Exa's open-sourced company search benchmark, 605 retrieval queries, and added per-query timing to Exa's test setup. Precision at 10 is the share of the top 10 results that are correct. The reported results:
| Measure | Seltz | Exa | Parallel |
|---|---:|---:|---:|
| Top-1 accuracy | 0.63 | **0.68** | not reported |
| Precision at 10 | 0.53 | 0.49 | **0.62** |
| Top-10 hit rate | 0.86 | not reported | not reported |
| Median latency | **327ms** | about 1,100ms (3.4x Seltz's median, per Seltz) | 5,275ms |
The fair reading: Exa built the benchmark and leads on getting the single best answer first. Parallel posts the highest precision, at a much higher latency. Of the three, Seltz is the only one under a second, with higher precision at 10 than Exa (0.53 to 0.49). Two caveats apply. The post is a vendor running a competitor's benchmark, and Seltz returns full documents while others return snippets, so the outputs aren't identical. Test on your own queries before you choose.
## When should you use Seltz, and when a live search results page?
These answer different questions, and many teams use both.
| Your question | Better fit | Why |
|---|---|---|
| "What does the web say about X?" inside an agent loop | Seltz Search or Answer | Fast, cited, full text from an owned index |
| "Find companies matching these criteria" | Seltz `companies` scope | Structured, one company per result |
| "Tell me when something new appears about Y" | Seltz Monitor | Returns only new records on a schedule |
| "What does Google show a searcher in Chicago today, AI Overview included?" | A live SERP from that location | The answer is the results page itself, which varies by place and time |
| "How does our brand rank across 10 countries?" | A live SERP per country | Ranking and SERP features change by location |
Massive's [Search endpoint](https://www.joinmassive.com/web-search) covers the last two rows. It returns the live Google results page for a query as rendered HTML from a chosen country, region, or city, and can wait up to a minute for the AI Overview or "People also ask" block to load ([Massive Docs](https://docs.joinmassive.com/web-render/search)). Our earlier [comparison of web search APIs for AI agents](https://www.joinmassive.com/blog/web-search-apis-for-ai-agents-compared-seltz-exa-brave-and-render-apis) lays out more of these trade-offs. An agent that needs fast background knowledge calls Seltz. One that needs to know what a local searcher sees calls a live SERP (search engine results page). Our guide to [getting Google AI Overviews into your search results](https://www.joinmassive.com/blog/search-ai-overviews-paa) shows that side in practice.
## Frequently Asked Questions
### What is the best web search API for LLMs?
It depends on the job. Indexed APIs such as Seltz return fast, citable full text for agent loops. Others lead on different measures, for example Exa on top-1 accuracy in Exa's company benchmark. When the question is what a searcher in a specific place sees, use a live results page from that location.
### Does Seltz have a free tier?
Yes. New accounts receive $20 of free requests, according to Seltz's pricing page. After that, usage is pay as you go, starting at $3 per 1,000 searches on the `base` tier.
### Does Seltz use its own index?
Yes. Seltz's docs say results are "served from Seltz's own index rather than a live scrape," and its Search concepts page calls it an "owned index" that is updated frequently.
### How do I add Seltz to Claude or Cursor?
Use Seltz's hosted MCP server at mcp.seltz.ai/mcp. In Claude Code, add it with the `claude mcp add` command (HTTP transport), then authenticate with your API key through `/mcp`. Cursor and VS Code take the same URL in their MCP config.
### How fast is Seltz?
Seltz reports a 327ms median latency in its company search benchmark, and its homepage says it retrieves full web pages in under 200ms. The `base` search tier trades some ranking precision for speed.
## The bottom line
- Seltz is a web search API for LLMs that returns full-text, cited documents from its own index at roughly 300ms, with clear, per-search pricing.
- Its benchmark shows a real latency edge and competitive precision. Other providers lead on some measures, so run your own queries.
- Pair it with a live, location-specific SERP when the question is what searchers actually see.
To try both sides, start with Seltz's $20 free credit and Massive's [Search endpoint docs](https://docs.joinmassive.com/web-render/search).
## Sources
- Seltz, [homepage](https://www.seltz.ai/)
- Seltz Docs, [llms.txt](https://docs.seltz.ai/llms.txt), [Choose a search tier](https://docs.seltz.ai/search/guides/choose-a-tier), [Fetch concepts](https://docs.seltz.ai/fetch/concepts), [Search concepts](https://docs.seltz.ai/search/search-concepts), [Data concepts](https://docs.seltz.ai/data/data-concepts), [Pricing](https://docs.seltz.ai/pricing), [MCP Server](https://docs.seltz.ai/integrations/mcp)
- Seltz, Python SDK `seltz` 1.17.0 on PyPI
- Seltz, ["Company search benchmark"](https://seltz.ai/blog/company-search-benchmark)
- Dealroom, ["Seltz raises $12.5M seed to rebuild web search for AI agents"](https://dealroom.co/news/135823-seltz-raises-12-5m-seed-to-rebuild-web-search-for-ai-agents/)
- Massive Docs, ["Search"](https://docs.joinmassive.com/web-render/search)