Which search APIs with a free tariff should I connect to Claude Code?
We compare 15 services with the API: who has a free volume replenished every month, who is issued once, where the card is needed, which MCP endpoints work without installation and without a key. Data for September 2026.
Briefly
- As of September 2026, twelve of the fifteen services compared have a free volume with the API: for eight, it is updated every month (including Context7, which is searched only for documentation), for four it is issued once. Tavily (1000 credits per month), Firecrawl (1000 credits), Bright Data (5000 MCP requests) and Exa ($10 per month; that the card is not needed is known only by retelling, the vendor does not explicitly write this) are suitable for permanent work without a card.
- Serper (2500 requests), SearchAPI.io (100), Jina (10 million tokens) and You.com ($100) give out a free volume once. This is enough to try the service, but not for permanent work.
- Since February 2026, Brave requires a card on any plan: $5 credits per month, money is debited over them. Perplexity and Kagi do not have a free tariff.
- Almost all vendors keep a ready-made remote MCP endpoint, so you don't need to install anything. Exa, Firecrawl, Parallel Search MCP, Context7 and You.com (
?profile=free) partially respond even without registration. - Tavily, SerpAPI, Linkup and Bright Data accept the key directly in the address, and then it is stored in plain text in the Claude Code configuration. Where possible, pass the key through the header or OAuth.
This is the second of five articles in the series "Free Internet Search for Claude Code". In the series: built-in WebSearch and WebFetch and MCP connection; services with API and free tariff (this article); search without keys and on your server: SearXNG, DuckDuckGo, multi-slide servers; browsers, scrapers, workarounds and services that have stopped working; search from Russia, security, quality, token consumption and a cheat sheet "what to choose".
Claude Code is an Anthropic console AI agent for development that can search the Internet. The built-in search cannot be replaced, but you can connect another one next to it via MCP. MCP (Model Context Protocol) is an open protocol for connecting external tools to an agent. An MCP server is a program or remote service that, using this protocol, gives an agent new tools, such as searching or reading pages.
All the figures below are taken from the vendor pages and checked on 09/17/2026, that is, these are the data of the vendors themselves, unless otherwise specified. Some of the price pages are read automatically, so you should double-check before paying the price.
Which search APIs give you a free tariff in 2026?
Of the fifteen services in the tables below, twelve give free volume: eight replenish it every month, four are issued once. The remaining three (Brave, Perplexity, Kagi) do not have free volume without a card.
There are two types of free volume. The free tariff with a monthly top-up is updated every month, and you can work on it all the time. A one—time grant is a starting amount that is issued once upon registration, in fact, a trial period.
A credit is an internal payment unit of the service: different operations write off different numbers of credits. Therefore, "1000 credits" does not always mean 1000 searches, see the conditions column.
Free rate with monthly top-up
| Service | Free per month | Conditions and restrictions | Map | For a fee |
|---|---|---|---|---|
| Tavily | 1000 credits, update on the 1st | Basic search (basic) costs 1 credit, advanced — 2 (Tavily documentation) | Not needed | From $0.008 per loan |
| Firecrawl | 1000 credits | The search costs 2 credits for 10 results, the limit is 10 requests per minute; some tools work without a key | Not needed | Hobby — $16 per month when paying for a year |
| Exa | $10 credits; $20 when registering (about 2,800 searches) | The conditions are confirmed by the Exa documentation; it also responds without a key, with limits. That the card is not needed — according to the retelling, this is not on the vendor's pages | Not needed (not confirmed) | Search — $7 per 1000 requests |
| Bright Data | 5000 MCP requests, updated on the 1st | The volume is specified in the server README | Not needed | Couldn't find out |
| Apify | $5 platform credits | The unused balance is burned at the end of the month | Not needed | Starter — $19 per month |
| Linkup | Top-up to $20, about 4000 standard searches | Only for work email; standard search — $0.005 | Not found out | $0.005 per search |
| SerpAPI | 250 searches | No more than 50 searches per hour | Not found out | Starter — $25 per month for 1000 searches |
| Context7 | 1000 calls, after exhausting +20 per day | Only library documentation, no web search; endpoint works without a key | Not needed | Pro — $10 per seat |
There is a discrepancy on the Linkup, but it is understandable. The Linkup marketing page promises "4000 free requests", and this converges with the calculation of $20 - $0.005. Older reviews call €5 per month, probably this figure is outdated.
One-time grant
| Service | Free of charge once | Conditions and restrictions | Map | For a fee |
|---|---|---|---|---|
| Serper | 2500 requests (by secondary source) | There is no official MCP server, only community servers | Not needed | About $50 for 50,000 requests (according to ColdIQ, a secondary source) |
| SearchAPI.io | 100 requests | The SearchAPI MCP server could not be found | Not needed | $40 per month for 10,000 requests |
| Jina (Reader and Search) | 10 million tokens (according to Markaicode, secondary source) | The read_url tool | Not found out | Not verified: the pricing page returns a 404 error |
| You.com | $100 at registration | Without a key — 100 requests per day via ?profile=free (according to the search snippet, not confirmed by the original source) | Not found out | Not verified: the pricing page returns error 403 |
For a fee: there is no free volume without a card
| Service | What is free | Conditions | Price |
|---|---|---|---|
| Brave Search API | $5 credits per month, about 1000 requests | The card is required, there are write-offs in excess of credits; up to 50 requests per second | $5 per 1000 requests |
| Perplexity | There is no free rate | The fate of the monthly $5 for Pro subscribers is not clear (see below) | Search API — $5 per 1000 requests |
| Kagi | There is no free rate | — | $12 per 1000 requests |
According to Perplexity, the issue has not been resolved. Third-party sources write that the monthly $5 for Pro subscribers was quietly removed at the beginning of 2026 (goosed.ie , secondary source). The snippet of the official FAQ in the search results still promises these credits, but the page itself gives 404.
Which free search API is suitable for permanent work?
Tavily, Exa, Firecrawl and Bright Data are suitable for permanent free work: they have a monthly limit and do not need a card (for Exa, according to the retelling, the vendor obviously does not write this). Brave gives a comparable amount, but only with a card and with the risk of write-offs.
Serper, SearchAPI.io , Jina and You.com with a key, they are suitable to try the service, but not for permanent work: the grant will end, and there will be no replenishment. Apify, Linkup and SerpAPI are also replenished monthly, but the first one has a small volume ($5), Linkup needs a working email, and SerpAPI has only 250 searches per month.
Context7 stands alone: it searches not on the Internet, but on library documentation. It is convenient to keep it next to a web search when an agent writes code with an unfamiliar framework.
Do I need to install something to enable the search API?
Most often not: most vendors already have a hosted HTTP MCP server, and Claude Code connects to it at the address. A hosting (remote) MCP server is an MCP server that works on the vendor's side and is accessible via HTTPS. The local package, on the contrary, is launched on your machine via npx (Node.js) or uvx (Python).
For Windows, the remote option is especially convenient: you don't need Node, Python, and fiddling with npx. In the table, <key> and <token> are places for your key. Bearer is a way to pass the key in the HTTP Authorization header. OAuth means logging in via the browser without manually copying the key.
| Vendor | Deleted endpoint | Local package | Key tools |
|---|---|---|---|
| Tavily | https://mcp.tavily.com/mcp/?tavilyApiKey =<key> (or Bearer, OAuth) | npx -y tavily-mcp@latest | search, extract, map, crawl |
| Exa | https://mcp.exa.ai/mcp (works without a key, with limits) | exa-mcp-server | web_search_exa, web_fetch_exa |
| Jina | https://mcp.jina.ai/v1 | — | search_web, read_url, search_arxiv and others, about 21 tools |
| Firecrawl | https://mcp.firecrawl.dev/v2/mcp (as well as /v2/mcp-oauth, /v2/mcp-search) | npx -y firecrawl-mcp | scrape, search, map, crawl, agent — 25 tools in total |
| SerpAPI | https://mcp.serpapi.com/<KEY>/mcp | Python 3.13+ and uv | search (Google, Bing, YouTube and others) |
| Perplexity | https://api.perplexity.ai/mcp | npx -y @perplexity-ai/mcp-server | search, ask, research, reason |
| Linkup | https://mcp.linkup.so/mcp?apiKey=<key> | npx -y linkup-mcp-server | search, fetch, research |
| You.com | https://api.you.com/mcp (?profile=free — without a key) | — | you-search; with the key also answer, contents, research |
| Kagi | https://mcp.kagi.com/mcp (Bearer) | uvx kagimcp | kagi_search_fetch, kagi_extract |
| Bright Data | https://mcp.brightdata.com/mcp?token=<token> | npx @brightdata/mcp | about 69 instruments, divided into groups |
| Apify | https://mcp.apify.com (OAuth, Bearer) | npx @apify/actors-mcp-server | rag-web-browser, web-fetch, actors |
| Context7 | https://mcp.context7.com/mcp (without a key) | — | resolve-library-id, query-docs |
| Brave | There is no hosting | npx -y @brave/brave-search-mcp-server | web, news, image, video, brave_llm_context |
The table data is available from the README repositories: tavily-mcp, exa-mcp-server, jina-ai/MCP, firecrawl-mcp-server, serpapi-mcp, Perplexity MCP, linkup-mcp-server, kagimcp, brightdata-mcp, apify-mcp-server, context7, brave-search-mcp-server.
Brave has limitations even on paid plans. Local search (for businesses nearby) requires a Pro plan, without it the server rolls back to regular web search; additional fragments of extra_snippets pages are also available only on Pro (README Brave). Serper does not have an official MCP server, there are community servers garylab/serper-mcp-server and marcopesani/mcp-server-serper.
How do I connect the search API to Claude Code?
With one command, claude mcp add: for a remote server with the --transport http flag and address, for a local server with --transport stdio and the launch command after --. The --scope (or -s) flag sets the scope: user — for all your projects, project — for the current project via the .mcp.json file.
The commands below are taken from the vendors' README, with two caveats. The line claude mcp add for Exa is our output according to the standard syntax: the README shows only the installation by the plugin. The string for Brave is compiled according to the documented syntax of the Claude Code and the package name from the Brave README.
Tavily
The key is transmitted in the address, read the warning below.
claude mcp add --transport http tavily "https://mcp.tavily.com/mcp/?tavilyApiKey=<your-api-key>"Exa
The first line is the official plugin, the second is a direct connection to the HTTP endpoint (our output is based on the standard syntax).
claude plugin install exa@claude-plugins-official
claude mcp add --transport http exa https://mcp.exa.ai/mcpJina
claude mcp add -s user --transport http jina https://mcp.jina.ai/v1 --header "Authorization: Bearer ${JINA_API_KEY}"Firecrawl
The first line is a remote server without a key (keyless, according to the Firecrawl blog), the second is a local server with a key.
claude mcp add --transport http firecrawl https://mcp.firecrawl.dev/v2/mcp
claude mcp add --transport stdio --env FIRECRAWL_API_KEY=fc-... firecrawl -- npx -y firecrawl-mcpSerpAPI
claude mcp add --transport http serpapi https://mcp.serpapi.com/YOUR_SERPAPI_API_KEY/mcpPerplexity
claude mcp add perplexity --env PERPLEXITY_API_KEY="..." -- npx -y @perplexity-ai/mcp-serverBright Data
claude mcp add --transport http brightdata "https://mcp.brightdata.com/mcp?token=YOUR_API_TOKEN"Brave
Only locally, the card must be linked to the Brave account.
claude mcp add --scope user --transport stdio --env BRAVE_API_KEY=xxx brave-search -- npx -y @brave/brave-search-mcp-serverThe .mcp.json file with the project area usually ends up in the repository, and the key from the address will go along with the code to everyone who has access. If agents work with the company's code and data, it is better to fix the rules for storing keys right away; this is part of the information security work.
Is it possible to search without registration and an API key?
Yes: several hosting MCP servers respond without registration at all, but with limitations in terms of volume or set of tools. Keyless endpoint is a remote MCP server that does not need either an account or a key.
| Service | Address or method | What is available without a key | Restrictions |
|---|---|---|---|
| Parallel Search MCP | https://search.parallel.ai/mcp | Free search without account and key, launched in September 2026 | Not specified |
| Exa | https://mcp.exa.ai/mcp | The server responds to requests without a key | There are limits, but they are not documented |
| You.com | https://api.you.com/mcp?profile=free | Search-only profile (based on the search snippet, not confirmed by the original source) | 100 requests per day |
| Firecrawl keyless | https://mcp.firecrawl.dev/v2/mcp | Only firecrawl_scrape, firecrawl_search and firecrawl_parse (by README) | The vendor promises 1000 free credits per month |
| Jina Reader | The read_url tool | Reading pages, without searching | 20 requests per minute |
| Context7 | https://mcp.context7.com/mcp | Library documentation search, not web search | 1000 calls per month, then +20 per day |
The launch date of the keyless Firecrawl mode differs in the sources: 06/16/2026 or restart 08/28/2026. Parallel Search MCP is enabled like this:
claude mcp add --transport http --scope project parallel-search https://search.parallel.ai/mcpThese services have two common limitations. The vendor can change the terms at any time, and your search queries go to a third party. If requests can reveal commercial information, it is more reliable to search on your server — the next article in the series is about it.
What should a company that connects search agents consider?
For one developer, a free tariff is usually enough, for the team it is more important who pays, where the keys are stored and where the requests go. If agents work for several employees, a shared key with a monthly limit will end faster than it seems: in a research task, the agent calls the search multiple times.
Any external search brings untrusted text from someone else's pages into the agent's context. Therefore, the source of the MCP server and the permissions granted to it mean no less than its quota. We select and configure such tools when implementing AI agents and when developing in the customer's circuit.
Products where data retrieval and collection are already configured
If you need a result, not an API choice: ready-made solutions that monitor sites and respond to your knowledge base.
Tracking changes on websites
The service monitors the necessary pages of competitors and partners and sends a Telegram notification with exactly what has changed on them. Changes are no longer noticed by accident.
Request a KPAI-Knowledge base and consultant
The AI itself collects the "question — answer" database from your site and from the history of work chats, updates it when pages change, and responds to customers using this database. If necessary, the answer first goes to the employee for approval.
Request a KPAI seller in Telegram
The AI bot guides the buyer through the stages of the sale: finds out the task, budget and necessary parameters, selects products from your catalog and gives them to the client's manager, who already understands what he is buying. The stages and rules of communication are configured without a programmer.
Request a KPHow we do it
Audit of tasks and restrictions
We find out what tasks agents solve with the search, through which provider the model works, whether foreign services are available from your network and what data can be included in search queries.
Search requirements: volume, availability, valid data
Search selection and limit calculation
We compare cloud APIs by quotas and payment terms, or we offer our own SearXNG. We check the selected services based on your real requests and evaluate whether there is enough free volume for the team.
Basic and backup search with an understandable cost
Deployment in your contour
We connect remote endpoints or deploy MCP servers and SearXNG in Docker on your server. Setting up the Claude Code configuration for a command without keys in addresses and repositories.
Search works the same for all employees
Rights and protection against prompt injection
We restrict the permissions of the MCP servers, add keys to environment variables and secrets, and set up rules that prevent the agent from following instructions from the found pages.
The agent searches, but does not act on someone else's instructions
Support
We monitor changes in tariffs and endpoints, spending limits and server updates, and switch to backup search if the vendor changes the conditions.
The search does not disappear after the next tariff change
We will select the search for your task and budget
Describe how many requests agents make and what are the payment and data restrictions. Let's calculate the options and set up the connection.
- We will respond within a working day
- Let's clarify the task and limitations
- We will offer a solution and pilot terms
Is it more convenient in the messenger?
+79262103289Employees respond from 9:00 to 23:00 Moscow time, every day
How can we help
AI agents
Not a demo chatbot, but an assistant on your data: from 1C, Access Control systems, CRM and regulations. Responds to guests, prompts employees and undertakes typical operations.
YandexGPTGigaChat1ССКУД
Development
When a typical program does not fit the process, and the tables no longer cope. We do it on ready-made platforms where it is reasonable, and we write from scratch where it is needed.
Веб-приложенияREST APIЛичные кабинеты1С
Information security
We find where the business is vulnerable, we consider what the incident will cost, and we design protection that does not interfere with work.
152-ФЗСетевое оборудованиеСерверы и СХД1СFrequent questions
Which search API can I use for free all the time?
As of September 2026, Tavily (1000 credits), Exa ($10 credits; the fact that the card is not needed is not confirmed by the original source), Firecrawl (1000 credits) and Bright Data (5000 MCP requests) provide monthly replenished free volume without linking the card. Apify ($5), Linkup (up to $20, only for work email) and SerpAPI (250 searches) are also replenished. Check the conditions on the pricing page before signing up.
Does Brave Search API have a free tariff in 2026?
There is no longer a free plan without a card: according to secondary sources, since February 2026, every Brave plan requires a card. Brave gives $5 credits per month, which is about 1000 requests, and in addition to them, it charges $5 per 1000 requests. Brave does not have a hosting MCP server, only a local package.
Is it possible to connect a web search to Claude Code without an API key?
Yeah. Exa remote MCP servers respond without registration (https://mcp.exa.ai/mcp ), Parallel Search MCP (https://search.parallel.ai/mcp ), Firecrawl in keyless mode (three tools) and You.com with ?profile=free (100 requests per day, not confirmed by the original source). Their limits are limited, and conditions can change at any time.
How does a one-time grant differ from a free tariff?
The free tariff is updated every month, and you can work on it all the time. A one-time grant is issued once upon registration, after it is exhausted, the service becomes paid. One-time grants for September 2026 give Serper (2500 requests), SearchAPI.io (100), Jina (10 million tokens) and You.com ($100).
How to connect Tavily or Exa to Claude Code?
With one claude mcp add command with the --transport http flag and the endpoint address: for Exa, this is https://mcp.exa.ai/mcp , for Tavily — https://mcp.tavily.com/mcp / with a key. Exa also has an official plugin: claude plugin install exa@claude-plugins-official. Set the Node.You don't need js or Python for remote endpoints.
Is it safe to pass the API key in the address of the MCP server?
Not the best option: the key in the address is stored in plain text in .claude.json or .mcp.json, and the second file often ends up in the repository. This is how Tavily, SerpAPI, Linkup, and Bright Data accept the key. If the vendor supports the Authorization or OAuth header, use them and keep the key in the environment variable.
Do Perplexity and Kagi have a free API?
No. The Perplexity search API costs $5 per 1000 queries, Kagi costs $12 per 1000. Third-party sources say that the monthly $5 for Perplexity Pro subscribers was removed at the beginning of 2026, but this has not been officially confirmed.
Sources
All sources are open and verified on 09/17/2026. Vendors change prices and limits without warning, check the rates page before paying.
- Tavily: prices — tavily.com , verified on 09/17/2026
- Tavily: API credits — docs.tavily.com , verified on 09/17/2026
- Firecrawl: prices — firecrawl.dev, verified 09/17/2026
- Exa: prices — exa.ai , verified on 09/17/2026
- Exa: Billing and Rate Limits — exa.ai , verified on 09/17/2026
- Bright Data: MCP server — brightdata.com , verified on 09/17/2026
- brightdata-mcp (README) — github.com , verified on 09/17/2026
- Apify: prices — apify.com , verified on 09/17/2026
- Linkup: price documentation — docs.linkup.so , verified on 09/17/2026
- Linkup: prices — linkup.so , verified on 09/17/2026
- SerpAPI: prices — serpapi.com , verified on 09/17/2026
- Context7: plans — context7.com , verified on 09/17/2026
- Brave Search API — brave.com , verified on 09/17/2026
- Serper — serper.dev, verified on 09/17/2026
- ColdIQ: Serper pricing (secondary source) — coldiq.com , verified on 09/17/2026
- SearchAPI.io : prices — searchapi.io , verified on 09/17/2026
- Jina Reader — jina.ai , verified on 09/17/2026
- Markaicode: Jina AI pricing (secondary source) — markaicode.com , verified on 09/17/2026
- You.com : MCP server — you.com , verified on 09/17/2026
- Perplexity: API prices — docs.perplexity.ai , verified on 09/17/2026
- Perplexity Pro quietly removes free API credits (secondary source) — goosed.ie , verified on 09/17/2026
- Kagi: API prices — kagi.com , verified on 09/17/2026
- Brave drops free Search API tier (secondary source) — implicator.ai , verified on 09/17/2026
- tavily-mcp (README) — github.com , verified on 09/17/2026
- exa-mcp-server (README) — github.com , verified on 09/17/2026
- jina-ai/MCP (README) — github.com , verified on 09/17/2026
- firecrawl-mcp-server (README) — github.com , verified on 09/17/2026
- serpapi-mcp (README) — github.com , verified on 09/17/2026
- Perplexity MCP (README) — github.com , verified on 09/17/2026
- linkup-mcp-server (README) — github.com , verified on 09/17/2026
- kagimcp (README) — github.com , verified on 09/17/2026
- apify-mcp-server (README) — github.com , verified on 09/17/2026
- context7 (README) — github.com , verified on 09/17/2026
- brave-search-mcp-server (README) — github.com , verified on 09/17/2026
- garylab/serper-mcp-server — github.com , verified on 09/17/2026
- marcopesani/mcp-server-serper — github.com , verified on 09/17/2026
- Firecrawl: keyless launch — firecrawl.dev, verified 09/17/2026
- parallel-web/search-mcp (README) — github.com , verified on 09/17/2026
Other articles in the series
Search in Claude Code: WebSearch, WebFetch and MCP
We analyze the built-in WebSearch and WebFetch tools in Claude Code: what they return, how much they cost, where they don't work and why the search engine can't be replaced in them. And we show you how to connect a search MCP server with one command and not step on a rake with scope, keys and permissions.
Digital Employee: what the bot will take from the administrator
The administrator of a fitness club or thermal spa complex answers the same questions all day long. We are sorting out what can be given to a digital employee for September 2026, what to leave to people, in which messenger to launch the bot and how to accept payment in it according to the law.
Private AI: customer data and someone else's cloud
Employees are increasingly inserting customer uploads into ChatGPT and Claude. Let's look at what 152-FZ says about this, what fines are in effect from May 30, 2025, and how to use AI without sending data to someone else's cloud.