What is Tavily? The AI search API, explained
Tavily is a web search and extraction API built for AI agents and RAG, with an MCP server, Python and TypeScript SDKs, and a free tier.
Tavily is a search API built for AI agents and retrieval-augmented generation (RAG). It gives a model one secure endpoint to search the live web, pull clean page content, crawl sites, and run multi-step research tasks. On StackResolve it scores 86 out of 100 for agent readiness, with a perfect 100 on Understanding.
#What Tavily does
Tavily calls itself the web access layer for AI agents. Instead of scraping raw HTML and cleaning it yourself, you send a query and get back ranked, LLM-ready results. The API covers five jobs:
| Endpoint | What it returns |
|---|---|
| Search | Ranked web results with short content snippets |
| Extract | Clean text pulled from specific URLs |
| Crawl | Content across many pages of a site |
| Map | A list of URLs under a site, no page content |
| Research | A multi-step research answer over several sources |
Under the hood Tavily runs real-time search, intelligent caching, PII detection, prompt injection protection, and content validation. The point is to hand a model text it can trust, not a page it has to parse.
#When to use it
Reach for Tavily when an agent needs facts the model was not trained on: today's news, a competitor's live pricing, a fresh docs page, or anything past the training cutoff. RAG pipelines use it to ground answers in current sources. Research agents use the /research endpoint to chase a question across several pages and return a synthesized answer.
If your agent only needs one known page, Extract is cheaper than a full search. If it needs to discover what pages exist, Map returns URLs without spending credits on content. For a side-by-side on ranking quality and pricing, see Tavily vs Exa. Both sit in the web search category on the registry.
#How an agent connects
Authentication is a single API key sent as a Bearer token. An agent signs up, gets a key, and starts calling within minutes. Three paths in:
- MCP server. Tavily publishes a Model Context Protocol server, so agents in Claude, Cursor, and Codex can load its tools directly. No wrapper code.
- SDKs. Official Python (
tavily-python) and TypeScript (@tavily/core) SDKs handle auth and response parsing. - REST plus OpenAPI. A documented OpenAPI spec and a published
llms.txtmean an agent can read the contract and build a call on its own.
A basic Python search takes three lines: import the client, pass your key, call search(). The response comes back as structured JSON with a stable shape, which is why Tavily scores 88 on Operability.
#The AgentReady score
Tavily's 86 breaks down like this:
| Pillar | Score |
|---|---|
| Discovery | 87 |
| Understanding | 100 |
| Adoption | 70 |
| Operability | 88 |
Understanding is perfect: every capability is defined, the API reference is structured, errors and limits are documented, and request and response examples are published. Operability is strong because output is predictable and rate-limit behavior is exposed in headers.
Adoption is the soft spot at 70. Tavily passes on self-service signup, programmatic credentials, a free allowance, both SDKs, a CLI, and MCP. It loses points for no copyable quickstart and a signup flow an agent cannot fully complete on its own. StackResolve's two open fixes: add a copyable quickstart so first success is fast, and support an Idempotency-Key header so retries are safe.
#Pricing and free tier
The Researcher tier is free and includes 1,000 API credits per month. After that, Pay As You Go charges $0.008 per credit, so you pay only for calls you make. Enterprise pricing is custom. The free tier is enough to prototype a RAG agent before spending a dollar.
Tavily raised a $25M Series A in 2025 and has since announced it is joining Nebius. It powers real-time search inside IBM WatsonX, the Databricks MCP marketplace, and JetBrains tools, and it ranked #1 on the SealQA and SimpleQA benchmarks. Full signal-by-signal detail lives on the Tavily registry profile.