Tavily

An agent gets read access to the public web as tools (search, extract, crawl, map and research), called through Tavily's API, SDKs or a remote MCP server. That is one integration with read access only. It takes no actions in other systems, and there are no credentials per user to scope, revoke or rotate.

Sourcedocs.tavily.com/llms.txtread 2026-09-22

Workflow logic lives in the customer's own framework, along with the customer's agent. Tavily documents no workflow runtime in which the customer can sequence, branch, retry or route steps, or combine deterministic workflow nodes with autonomous agent steps. Its Research endpoint runs a fixed pipeline of its own, planning, searching and synthesizing a cited report, and the customer cannot define, branch or sequence steps in it.

Sourcedocs.tavily.com/documentation/api-reference/endpoint/researchread 2026-09-22

Searches and extraction run against the public web. Tavily maintains no retrieval structure over the customer's own documents or data, and no ingestion, retrieval or RAG layer grounds agents in company data; an index of the public web is not the customer's corpus. Its Crawl to RAG and hybrid-research examples show customers building their own knowledge base with other tools.

Sourcedocs.tavily.com/llms.txtread 2026-09-22

The service retrieves public web content and takes no actions in other systems. No approval step, consent checkpoint, review queue, escalation rule, guardrail, pause or resume control, or policy constraint on what an agent does with that content is documented. API keys carry credit limits and expiry, which are account controls rather than oversight of agent actions.

Sourcedocs.tavily.com/llms.txtread 2026-09-22

Access controls are documented. Enterprise customers generate and deactivate API keys programmatically, with admin or user permission levels, scopes, development or production types, credit limits per key and expiry. Organization usage is reported per key, and the remote MCP server supports OAuth 2.0. The site footer links a Trust Center at trust.tavily.com, which loads only in a browser, so no compliance attestation is confirmed. No SSO is documented.

Sourcedocs.tavily.com/documentation/enterprise/generate-keysread 2026-09-22

On paid plans the Logs API returns usage logs per request for the account's or organization's keys, filterable by date and endpoint, and the Usage and Organization Usage APIs report credits and request counts per key. The Logs page states that logs never include the input or output of a request, so what the agent asked and received cannot be inspected.

Sourcedocs.tavily.com/documentation/api-reference/endpoint/logsread 2026-09-22

Each request is answered independently, with no agent context kept across calls; research tasks are retrieved by request ID until they finish. No memory layer with a stated scope and lifetime is documented.

Sourcedocs.tavily.com/llms.txtread 2026-09-22

Customers reach Tavily as a hosted API. It is listed on the Snowflake Marketplace as a native app installed into the customer's Snowflake account, and on Amazon Bedrock AgentCore, Azure AI Foundry and Databricks, but the Snowflake app is configured with a Tavily API key and calls Tavily's hosted service, so the search itself does not run in the customer's environment. No self hosted, private cloud or region option is documented.

Sourcedocs.tavily.com/documentation/partnerships/snowflakeread 2026-09-22

The Research endpoint is a prebuilt research agent that plans, searches, analyzes sources and returns a cited report, in mini, pro or auto modes, streamed or polled. Tavily also publishes Agent Skills for coding agents and an Agent Toolkit of reference research agents as code. These are working assets built for one job. No browsable catalog of prebuilt workflows, templates or agents for specific roles is published, and the Examples Hub is tutorial code.

Sourcedocs.tavily.com/documentation/api-reference/endpoint/researchread 2026-09-22

Work starts only when an agent or application calls Tavily. No schedule, webhook, event subscription or channel that starts an agent is documented, so as a request and response API it has no way to wake an agent on its own.

Sourcedocs.tavily.com/llms.txtread 2026-09-22

Tavily serves web data to whatever model the customer's agent runs: its docs carry integrations for OpenAI and Anthropic tool calling, Google ADK, the Vercel AI SDK, LangChain, LlamaIndex, CrewAI, Mastra and others, and any MCP client can connect, each with the customer's own model. Only its own Research endpoint picks a Tavily-run model tier (mini, pro or auto). Model choice for the customer's agent stays with the customer, who brings its own providers.

Sourcedocs.tavily.com/llms.txtread 2026-09-22

Tavily is callable from outside through a documented REST API with an OpenAPI specification (search, extract, crawl, map, research, usage, logs and Enterprise key management), Python and JavaScript SDKs, a CLI with Agent Skills, and a remote MCP server with OAuth 2.0, and it is embedded in dozens of agent frameworks and platforms.

Sourcedocs.tavily.com/documentation/api-reference/introductionread 2026-09-22

A Feedback endpoint (beta) collects feedback on search results to improve Tavily, and the RAG Evaluation example is tutorial code; neither tests or scores the customer's agent. Tavily's published benchmark results describe its own search. No testing, scoring or quality gates for the customer's agent workflows are documented.

Sourcedocs.tavily.com/llms.txtread 2026-09-22

Web pages that have no API can be read and traversed through Tavily's Extract, Crawl and Map endpoints, which return clean content; that is scraping. No interactive control of a real browser interface (clicking, typing, filling forms or operating a logged in session) is documented.

Sourcedocs.tavily.com/documentation/api-reference/endpoint/crawlread 2026-09-22