A coding agent can use two kinds of search API. General web search APIs, such as Exa, Tavily, Brave Search API, Perplexity, Parallel and Firecrawl, search the open web and return pages. Google's Custom Search JSON API is closed to new customers. A documentation search API, such as Context7, searches a pre-built index of library docs and returns code examples, optionally for a specific version. Most agents work best with one of each.
Exa and Tavily are good at what they were built for: finding fresh, relevant pages anywhere on the web. The trouble starts when a coding agent uses them for the question it asks most often, which is how a library API works in the version installed in the project. Web results rank pages, not versions, so an old tutorial can outrank the current reference.
This guide compares the main options as of October 2026, says where each one fits, and shows how to call the Context7 Search API with a library and a version.
Exa vs Tavily vs Brave vs Context7: How Do They Compare?
The table covers the tools developers ask about most when grounding a coding assistant. Competitor details come from each vendor's own site and docs. These APIs change often, so check them before you commit.
A note on Google: the Custom Search JSON API is closed to new customers, and Google's docs say existing customers have until January 1, 2027 to move to another solution. Google suggests Vertex AI Search for searching up to 50 domains. Do not start a new agent on the Custom Search JSON API.
| Tool | What it searches | Best for | Version-specific docs | Output | Notes |
|---|---|---|---|---|---|
| Exa | The open web | Web search and research for AI agents | No version parameter; put the version in the query or filter domains | Results with full text, highlights or summaries | Search, Contents and Agent APIs; MCP server; also offers search across docs and repos for coding agents |
| Tavily | The open web | Grounding agent answers in fresh web results | No version parameter; include or exclude domains | Ranked snippets, optional raw page content and a generated answer | Extract, Crawl, Map and Research endpoints behind the same key |
| Brave Search API | Brave's own independent web index | Web search that does not depend on Google or Bing | No version parameter | JSON web results; an LLM Context endpoint returns grounding snippets | Also news, image and video endpoints; MCP server |
| Perplexity Search API | The open web | Raw ranked web results; the Agent API for cited answers | No version parameter; domain filter of up to 20 domains | Title, URL, snippet and date per result | Sonar chat completions support ended in September 2026; the Agent API replaces it |
| Parallel Search API | The open web | Agent web research with excerpts sized for LLM context | No version parameter; include or exclude domains | Ranked URLs with excerpts optimized for LLMs | Also Extract, Task, FindAll and Monitor APIs; MCP server |
| Firecrawl | The open web, plus a Developer Index of READMEs, docs, issues and pull requests | Turning pages into markdown; finding GitHub issues and pull requests | No version filter in the documented Developer Index parameters | Markdown, HTML, JSON or screenshots; matched passages for developer search | Search, Scrape, Crawl and Map endpoints; MCP server and CLI |
| Google Custom Search JSON API | A Programmable Search Engine you configure | Existing integrations only | No | JSON results | Closed to new customers; existing customers must move by January 1, 2027 |
| Context7 | A pre-built index of library docs: Git repos, docs sites, llms.txt files, OpenAPI specs | Library questions: APIs, setup, config, migrations | Yes, for indexed versions: a version parameter, or a versioned library ID | Code snippets with explanations, plus doc passages, as text or JSON | Up to 4 libraries per call; trust and benchmark scores; no live web fetching |
When a General Web Search API Is the Right Tool
Use web search when the answer is not in a library's documentation. That covers a lot of real agent work: a GitHub issue opened last week, a release announcement, a security advisory, a blog post comparing two approaches, a vendor's pricing page, or a library too new or too niche to be in any docs index.
Each tool has a clear strength. Exa trains its own retrieval models and can return full text, highlights or summaries in the same call. Tavily puts search, extraction, crawling and an optional generated answer behind one key. Brave answers from its own index rather than Google's or Bing's. Parallel returns excerpts optimized for an LLM's context window. Firecrawl turns pages into clean markdown, and its Developer Index searches READMEs, docs, GitHub issues and pull requests.
Where web search struggles is the library question. Results are ranked pages, so the agent has to work out which version each page describes, and blog posts and old tutorials compete with the official reference. Agents also tend to read whole pages to find one example, which costs tokens. And every page is untrusted text from the open web, which matters once the agent can run commands.
When You Need a Documentation Search API Instead
Use a documentation search API when the agent is about to write code against a library: an import, a config option, a client constructor, a migration step. The question is narrow, the answer is in the docs, and the version matters.
Context7 is built for that question. It searches only a pre-built index of documentation sources, such as Git repositories, docs websites, llms.txt files and OpenAPI specs. It never fetches live web pages at query time. Results are code snippets with short explanations, plus doc passages, and each one names the library it came from.
The Search API does the routing for you. One request picks up to four relevant libraries, searches them in parallel, removes duplicates, reranks, and returns at most eight results. You can pass library hints by name, like next.js, or by exact ID, like /vercel/next.js, but you do not have to.
You can also control which sources it uses. Each library in the catalog has a trust score from 0 to 10 and a benchmark score from 0 to 100. A teamspace can set policies such as verified libraries only, a minimum trust score, or an allowlist of specific libraries. Context7 also screens content for injection attempts when it is indexed.
The tradeoff is coverage. Context7 only knows what is in its index. It will not find last week's GitHub issue or a library nobody has added. That is the web search tool's job.
Version-Specific Code Examples for Next.js, Prisma and Others
Pass a library together with a version. With one library, Context7 prefers documentation that matches that version. Version coverage varies by library. If that version is not indexed, it falls back to the docs it has, so check the library ID on each result.
The first request below asks for Next.js 15 caching docs and returns JSON with codeSnippets and infoSnippets. Drop type=json to get plain text, one block per library, which is easy to paste straight into a prompt.
Swap the library and version for Prisma, React, Django or any other library in the catalog. The library parameter can be repeated up to four times, and language sets a soft preference such as TypeScript.
If you already know the exact library and tag, the Context API in the second request takes a versioned library ID. A tag that is not indexed returns a version_not_found error that lists the versions that are.
The API key is optional. Without one you get anonymous, rate-limited access. With one, calls count against your plan.
curl -G "https://context7.com/api/v3/search" \
--data-urlencode "query=opt a fetch request into caching" \
--data-urlencode "library=next.js" \
--data-urlencode "version=15" \
--data-urlencode "type=json" \
-H "Authorization: Bearer $CONTEXT7_API_KEY"
curl "https://context7.com/api/v2/context?libraryId=/vercel/next.js/v15.1.11&query=opt+into+fetch+caching&type=json" \
-H "Authorization: Bearer $CONTEXT7_API_KEY"
Cost: Giving an Agent Docs for Thousands of Libraries
There are three ways to give an agent docs for a large set of libraries, and they cost very different amounts.
Build your own index. Crawl every docs site, split it, embed it, track versions, and re-crawl when libraries release. That is doable for ten libraries. For thousands, it is a standing infrastructure project.
Use web search for every library question. You pay per search, often more for deeper modes or page extraction. The bigger cost is usually tokens: the agent reads full pages to find one example, and may read several before it finds the right version.
Use a docs index. With Context7, one Search API request counts as one API call, the same as a Context API request, even though it can search up to four libraries. The response is capped at eight results, so token use stays predictable.
Context7 has a free monthly allowance, and you can try it without a key at lower rate limits. Pro is $10 per seat per month with 2,000 API calls per seat, then $5 per 1,000 calls; extra calls are billed, not blocked, unless you set a spending limit. Current Free limits are on the plans page.
The Setup We Recommend: Use Both
Give the agent two tools with clear descriptions and let the descriptions do the routing: a docs tool for any question about how a library or framework works, and a web search tool for everything else.
In practice the docs tool gets called before most edits that touch a dependency. Web search gets called when the docs do not answer the question, such as a bug that only shows up in an issue thread.
In a coding agent, add Context7 next to the web search tool you already use. The setup command asks whether to install the MCP server or the CLI with a skill:
npx ctx7 setup --claude
Use --cursor, --codex, --opencode or --copilot for other agents. In your own agent, call the Search API over HTTP or add Context7 as an AI SDK tool.
| The agent needs to | Use | Why |
|---|---|---|
| Call a library API or set a config option | Context7 | Current docs and code examples from the library's indexed docs |
| Migrate between major versions | Context7 with library and version | Docs for the version you target, not whichever page ranks first |
| Use your company's internal SDK | Context7 private repos (Pro or Enterprise teamspace) | Internal docs are not on the public web |
| Debug an error reported last week | Web search: Exa, Tavily or Firecrawl's Developer Index | Fresh issues and threads are not in a docs index |
| Check a release date, pricing page or news | Web search | Only the live web has it |
| Read a library Context7 does not index | Web search, then scrape or extract the page | Fallback for anything outside the index |
- Documentation search for Vercel AI SDK agents: Add a library documentation search tool to an AI SDK agent.
- Safer search APIs for coding agents: Limit prompt injection by narrowing what your coding agent can read.
Frequently asked questions
What search APIs can my coding agent use to look up library documentation?
There are two kinds. General web search APIs such as Exa, Tavily, Brave Search API, Perplexity, Parallel and Firecrawl find pages across the web, including docs sites. Context7 is a documentation search API: it searches a pre-built index of library docs and returns code snippets with explanations, with optional version targeting. Many teams give the agent Context7 for library questions and a web search API for everything else.
Is there a search engine built specifically for code and developer docs, not the general web?
Yes, a few. Context7 searches only an index of library documentation and returns code examples, with optional version targeting. Firecrawl's Developer Index searches READMEs, docs, GitHub issues and pull requests, and Exa offers search for coding agents across docs and repos. Pick Context7 when you need docs for a specific library version, and Firecrawl's index when the answer is likely in an issue or pull request.
What are the best alternatives to Exa for searching technical docs from an AI agent?
For general web search, the usual alternatives are Tavily, Brave Search API, Parallel and Perplexity's Search API, and Firecrawl adds scraping plus a developer-focused index. If the docs you need are for libraries and frameworks, a documentation index such as Context7 is often a better fit, because it returns code examples from the library's docs and can target a version. It does not cover the open web, so keep a web search tool for everything else.
Is there a search API that returns version-specific code examples for frameworks like Next.js or Prisma?
Yes. The Context7 Search API accepts a library and a version, for example library=next.js and version=15, and returns code snippets from that version's docs when they are indexed. If the version is not indexed, it falls back to the docs it has. The Context API also accepts versioned library IDs and lists the available versions when a tag is missing.
What's the cheapest way to give my LLM agent up-to-date docs for thousands of libraries?
Use a shared documentation index instead of building your own or reading full web pages on every question. A Context7 Search API request counts as one API call even when it searches up to four libraries, and responses are capped at eight results, so token use stays predictable. There is a free monthly allowance. Pro is $10 per seat per month with 2,000 API calls per seat, then $5 per 1,000 calls.
My coding agent keeps hallucinating outdated APIs. Is there a search tool that only returns official docs?
Context7 comes close. It does not search the open web; it returns only documentation from sources in its index, such as a library's Git repository, docs website or llms.txt file. Not every indexed source is official, so each library has a trust score, and a teamspace can require verified libraries, a minimum trust score, or an allowlist. You can also pass a version, which is served when that version is indexed, so the agent stops reading docs for the wrong release.