Brave Search API
Also known as: Brave Search API, Brave API, Brave AI Grounding, Brave Answers API, Brave LLM Context, Brave Search MCP Server
The Brave Search API serves Brave's independent index of more than 40 billion pages to AI agents and developers. Its LLM Context endpoint returns extracted, ranked passages within a set token budget, and Answers returns cited, multi search answers through an OpenAI compatible interface, with an official MCP server for Claude and other clients. Queries are never stored or trained on, it holds SOC 2 Type II, and search costs $5 per 1,000 requests with $5 in free monthly credits.
The Brave Search API sells access to Brave's own web index, more than 40 billion pages built by its own crawler with no fallback to Google or Bing, to AI labs and developers building agents. The same index runs Brave Search, the consumer search engine of Brave Software, maker of the Brave browser, which handles hundreds of millions of queries a day and has answered 150 billion in all.
AWS, Cohere, Mistral AI, Snowflake, you.com, Firecrawl and Kagi are among the AI companies using the API, with Shopify, Chegg, Turnitin, Axel Springer, Motley Fool, Apollo, Causaly, SoftwareOne and Together.ai among other customers, and more than a million sign ups have come through the OpenClaw and Hermes agents.
Two endpoints are built for agents. LLM Context returns ranked, extracted passages from pages, including text, tables, code, JSON, forum threads and video captions, in one call with a token budget the developer sets from about 1,000 to 33,000 tokens, so an agent's model can read the web without a separate scraper.
Answers returns summarized answers with citations, entities and structured data, grounded on one or more searches, with a multi search research mode, streaming and an OpenAI compatible chat completions interface. Web, news, image, video and local place results with map data serve human facing search, and Goggles let a team rerank results by its own rules.
An official open source MCP server exposes web, local, image, video, news, summarizer and context search to Claude, Cursor and other compatible clients, and official skills help coding agents call it.
Zero data retention is built into the architecture rather than promised by contract: queries stay on infrastructure Brave runs end to end and are never shared or used for training, and storing results or training on them needs a custom plan. The service holds a SOC 2 Type II attestation and is sold directly and through AWS Marketplace. Search costs $5 per 1,000 requests at up to 50 queries a second, Answers costs $4 per 1,000 requests plus $5 per million input and output tokens at 2 queries a second, each plan carries $5 of free credits a month, billing runs on prepaid credits with no subscription, and Enterprise terms, capacity and custom endpoints are quoted.
The Brave Search API fits teams that need an independent, privacy preserving web index for agents and retrieval pipelines at a predictable per request price, especially where questions from users pass through the search and must not be kept. It is a data and answer service rather than an agent platform: it keeps no memory, starts no work on its own and leaves orchestration, approvals and the choice of model to the customer's agent.
Vendor details
Canonical URL
https://brave.com/search/api/
Category
Agent infrastructure
Subcategory
Independent web search index served to AI agents through LLM Context extraction, cited Answers and an official MCP server, with architectural zero data retention
Funding status
Product of Brave Software, Inc., a privately held company.
Company status
first party product
Use cases & customers
Primary use cases
Target customers
Deployment options
Integrations
REST API; OpenAI compatible chat completions endpoint for Answers; official open source MCP server for Claude, Cursor and other MCP clients; official skills for coding agents; AWS Marketplace; Snowflake partnership; works in any agent harness.
In practice
A coding assistant team wires the official MCP server into its agent, so when a developer asks about a library's new release the agent searches the web, pulls the changelog passage through LLM Context and answers with the source link.
A research startup calls the Answers endpoint in research mode for each analyst question, getting a cited summary built from several searches and the entities involved, and stores only the citations its own users approve.
An education company grounds its tutoring model in fresh web results through LLM Context with a 4,000 token budget per query, relying on zero data retention because student questions pass through the search.
Sources & related URLs
Related / legacy domains
Agentic Index coverage score
3.5 / 14 capabilities · 25%
| Integrations & Tool Calling | Partial |
|---|---|
|
Agents reach the index from any harness through the web API, an OpenAI compatible chat completions endpoint and the official MCP server, and the service is listed on AWS Marketplace, with Snowflake among its integration partners. It returns search data and answers to the calling agent and takes no action in any other system, such as a CRM or ticketing tool, on that agent's behalf. Sourcebrave.com/search/apiread 2026-11-19 |
|
| Workflow Orchestration | Not documented |
|
Each call is a single request and response. The Answers endpoint can run several searches before writing a cited answer, but it plans no multi step work across tools, keeps no workflow state between calls and hands nothing on to other agents, so any sequencing of searches, reading and actions sits in the customer's own agent or framework, such as LangGraph or a custom loop. Sourceapi-docs.search.brave.software/services/ai-groundingread 2026-11-19 |
|
| Knowledge Grounding & RAG | Not documented |
|
The product is the web index itself, more than 40 billion pages crawled by Brave, which a customer's agent uses as its source of current public information. It indexes no customer documents or private data, so grounding in a company's own policies, tickets or files has to come from another retrieval layer, with this service supplying only the open web half. Sourcebrave.com/search/apiread 2026-11-19 |
|
| Human Oversight & Guardrails | Not documented |
|
The service only reads the public web and returns results or answers, taking no action that a person would need to approve. There is no review step, policy layer or guardrail over what an agent does with what it finds, beyond Goggles and request parameters, such as safe search and freshness, that shape which results come back for a query and how they are ranked. Sourcebrave.com/search/apiread 2026-11-19 |
|
| Security, Identity & Governance | Full |
|
Brave holds a SOC 2 Type II attestation for the API, with a trust center for reviews, and builds zero data retention into the architecture: queries stay on infrastructure it operates end to end, are never shared with third parties and are never used for training, and storing results or training models on them requires a custom plan with its own terms. Sourcebrave.com/search/apiread 2026-11-19 |
|
| Observability & Auditability | Partial |
|
Answers come back with citations, entities and structured source data, so an agent or a reviewer can trace each claim to the pages behind it, and LLM Context passages carry their source metadata. Usage and spend are tracked against prepaid credits, but there is no log of an agent's queries over time to review afterwards, since queries are not retained. Sourceapi-docs.search.brave.software/services/ai-groundingread 2026-11-19 |
|
| Memory & State Persistence | Not documented |
|
Every request stands alone. With zero data retention built into the architecture, queries are not stored, so the service keeps no record of earlier searches, preferences or conversations to draw on later, and any memory of past research lives in the customer's own agent or database rather than with the search provider that answered each of the questions. Sourcebrave.com/search/apiread 2026-11-19 |
|
| Deployment & Data Residency | Not documented |
|
The API runs only as a hosted service on infrastructure Brave operates, reached over the internet and also sold through AWS Marketplace. There is no self hosted copy of the index, private deployment or regional endpoint a customer can choose, though zero data retention means query data is not kept wherever it is processed on the way to an answer or result. Sourcebrave.com/search/apiread 2026-11-19 |
|
| Prebuilt Agents / Templates / Packs | Partial |
|
Answers works as a ready research tool: it runs one or several searches, reads the results and returns a cited summary with entities, without the customer building that loop. Beyond it, the offering is a set of search endpoints, an MCP server and coding agent skills rather than finished agents for particular jobs that a team could deploy as they are. Sourceapi-docs.search.brave.software/services/ai-groundingread 2026-11-19 |
|
| Triggers & Channel Coverage | Not documented |
|
Nothing runs until a developer's code or agent calls the API. There are no schedules, monitors, webhooks or alerts that start a search when the web changes, and no messaging channel where a person can ask the service directly, so every query begins with a request from the customer's own application or from an MCP client acting for a user at that moment. Sourcebrave.com/search/apiread 2026-11-19 |
|
| Model Flexibility & Routing | Not documented |
|
Answers runs on a model the vendor operates, named simply brave in its chat completions interface, with no choice of model, provider or routing between them. Search results and LLM Context passages feed whatever model the customer's own agent uses, which is a choice the customer makes outside the service rather than one the service itself offers to make. Sourcebrave.com/search/apiread 2026-11-19 |
|
| APIs / SDKs / MCP Extensibility | Full |
|
The web API covers web, news, image, video, local, LLM Context and Answers endpoints, and Answers speaks the OpenAI chat completions format so existing SDK code works unchanged. Brave's official open source MCP server runs over standard input or the web and exposes the search tools to any compatible client, and official skills help coding agents call the API. Sourcebrave.com/blog/search-api-growthread 2026-11-19 |
|
| Testing, Debugging & Optimization | Not documented |
|
Speed and accuracy, with responses under 500 milliseconds at the 90th percentile, are properties of the search service rather than of the customer's agent. There is no tool for scoring answers against expected results, comparing Goggles or parameters side by side, or testing an agent's search behavior on a fixed question set before release, so that testing happens in the customer's own harness. Sourcebrave.com/search/apiread 2026-11-19 |
|
| Browser / Computer-use | Not documented |
|
LLM Context extracts the text of pages so agents can read them without a browser, but the service drives no browser or desktop session itself, so pages behind logins, forms or scripts that need clicking stay out of reach. Leo, the separate browsing agent built into the Brave browser, is where the company's computer use lives, for people browsing rather than developers. Sourcebrave.com/search/apiread 2026-11-19 |
|
The Agentic Index coverage score grades every vendor Full, Partial or Not documented against the same 14 buyer facing capabilities, from public evidence only. Each capability links to how all vendors in the index score on it. How this evidence is graded
Pricing
Search $5 per 1,000 requests; Answers $4 per 1,000 requests plus $5 per million tokens; $5 free credits a month per plan; Enterprise custom
Per 1,000 requests, plus input and output tokens on Answers, drawn from prepaid credits
Included quota
$5 of free credits a month per plan; Search up to 50 queries a second, Answers up to 2
What is public
Search costs $5 per 1,000 requests, and Answers costs $4 per 1,000 requests plus $5 per million input and output tokens. Each plan includes $5 of free credits a month, and Enterprise terms are quoted.
Billing mechanics
Billing draws on prepaid credits with no subscription or minimum, with Search capped at 50 queries a second and Answers at 2.
Cost watchouts
Answers adds $5 per million input and output tokens on top of the per request rate.
Variable cost rationale
Every search and every Answers token is billed as used, so an agent that searches often or reads long answers drives the cost directly.
Additional watchouts
The Answers rate limit of 2 queries a second is far below the Search limit.
Sales call required
Mixed (some tiers require a call)
Free / trial
$5 in free credits each month on each plan
Lowest paid plan
Search, $5 per 1,000 requests
Commercial notes
Storing results or training models on them requires a custom plan, and the API is also sold through AWS Marketplace.
Key ambiguities
The free monthly allowance appears as up to $20 in the headline offer and as $5 on each plan.
Cancellation / refund
No subscription or minimum.
Missing data
Enterprise pricing is not public.
Related vendors
- AgentOps — Agent observability and debugging platform: open source SDKs trace…
- AgentPhone — AgentPhone gives the AI agents a team builds their own US or…
- Agno — Python agent framework and AgentOS runtime (formerly Phidata) for…
- Airbyte — Airbyte moves data from operational systems into warehouses and…
- AIsa — Resource and payment gateway for AI agents, with one key to 110+…
- Albato — Albato is a no code integration platform that sells an embedded,…
Alternatives to Brave Search API
Other products that do the work Brave Search API does, and what each one is.
- NeverminedPayments infrastructure for AI agents
- Rime
- Skyfire
- SpeechmaticsSpeech to text and text to speech APIs for voice agents and transcription
- TavilyWeb access API for AI agents
- OpenBox AIRuntime governance layer