Official (Vendor)Active· 15d★ 4.3by Tavily

Tavily Search

Tavily's MCP — purpose-built search API for agents, with citations and structured snippets ready for synthesis.

View repo Homepage

Safety & Trust

Delv Safety Grade: A

Score 84/100 · assessed 2026-04-28

Maintainer75

Permissions95

Supply chain85

Transparency80

Incidents100

Tavily Search is a purpose-built search API for AI agents, offered as an official MCP server by Tavily, a recognised vendor in the agent search space. The server provides read-only search capabilities with structured citations, making it one of the safer search integrations available. It requires an API key (TAVILY_API_KEY) and makes outbound network calls to Tavily's service. The maintainer is a legitimate commercial entity focused on agent-specific search, though smaller than major cloud vendors. Supply chain is solid via npm with standard npx installation. Permissions are tightly scoped to search queries only, with no filesystem or shell access. Transparency is good with open source code and clear documentation. No known security incidents. The main consideration is the API key requirement and dependency on Tavily's external service, but this is inherent to the design and clearly documented.

Lethal Trifecta (prompt-injection exposure)

TWO OF THREE

Private dataNo

Reads secrets, credentials, private files

Untrusted inputYes

Ingests web pages, PRs, issues, emails

External commsYes

Can send data outbound

Same shape. Tavily caches pages it found, which means stale poisoned content can keep coming back.

Green flags

Read-only search API with no write or execution permissions
Purpose-built for AI agents with structured, citation-ready output
Official vendor package distributed via npm with standard install
Open source repository with clear documentation and examples
No known security incidents or credential leaks

Red flags

Requires API key stored in environment, potential credential exposure risk
Dependency on external Tavily service availability and pricing
Smaller vendor with less ecosystem scrutiny than major cloud providers

Permissions requested

Outbound networkAccess secrets

Assessed by Delv Editorial using public metadata. Grades are advisory and update as the ecosystem changes. They do not replace your own review of permissions and code before granting an agent access to sensitive systems.

Install

npx -y tavily-mcp

Env vars needed: TAVILY_API_KEY

Review

Tavily is a search API designed specifically for AI agents, and its MCP server makes that power available directly inside Claude and other supported hosts. Instead of generic web results, you get structured snippets with citations already formatted for synthesis. I've used it for research tasks where I need Claude to pull current information without hallucinating sources, and it's noticeably better than telling Claude to "search the web" through other tools. The key difference is intent. Tavily doesn't just scrape the top ten links. It returns answer-focused content with clear attribution, which means Claude can cite sources properly without you having to verify every claim manually. I'd reach for this when building research assistants, fact-checking workflows, or any agent that needs to reference current events or technical documentation that changes frequently. Setup is straightforward if you're on Claude Desktop: grab an API key from Tavily, drop it in your environment, and run the install command. The server exposes a single search tool that Claude can call whenever it needs external information. Results come back with URLs, snippets, and relevance scores. The quality depends on Tavily's indexing, which skews towards news, blogs, and public documentation. It's not a replacement for domain-specific databases or paywalled content. One quirk: Tavily is a paid service after the free tier, so you'll want to monitor usage if you're building something that searches frequently. The MCP server itself is just a thin wrapper around the API, which means performance and rate limits are determined by your Tavily plan, not the server. If you're already using Tavily's API elsewhere, this is a no-brainer. If you're new to it, the free tier is generous enough to evaluate properly. Who shouldn't bother: anyone who needs only local or private data, or developers who already have a search solution they trust. This is for people who want Claude to reference the open web reliably, with citations that actually check out.

Verdict

Install this if you need Claude to search the web with proper citations and you're willing to pay for quality results after the free tier. Skip it if your workflows are purely local or you already have a search API you prefer. It does one thing well and doesn't pretend otherwise.

Good at

Returns structured snippets with citations, not raw HTML or link dumps.
Purpose-built for agents, so results are already formatted for synthesis.
Works across multiple hosts including Claude Desktop, Cursor, and Zed.
Free tier is generous enough to evaluate properly before committing.

Watch out

Paid service after free tier, so costs scale with usage.
Quality depends on Tavily's indexing, which won't cover paywalled or niche sources.
Manual config required for hosts beyond Claude Desktop.
Single-purpose tool, no local search or document indexing features.

Getting started

1. Sign up at tavily.com and grab your API key from the dashboard. 2. Set the environment variable `TAVILY_API_KEY` in your shell or host config. 3. Run `npx -y tavily-mcp` to install, then add the server to your Claude Desktop or other host config file (check the repo for JSON examples). 4. Restart your host and ask Claude to search for something current, like recent news or a technical term. 5. Watch out for API usage if you're on the free tier, searches count against your quota quickly in long sessions.

Works with

Claude DesktopClaude CodeCursorWindsurfClineZed