Brave Search Skills - Setup Guide¶

Source: brave/brave-search-skills via skills.sh - 18.0K combined installs across 13 indexed listings; first documented November 4, 2026 (morning sweep) GitHub: 193 stars, MIT licence, last pushed 2026-10-29, 13 SKILL.md directories (12 under skills/, one under clawhub/) Category: Web search for agents: official Brave Search API skills for grounding, news, local and summarization workflows Quality Tier: 🟡 Yellow: bx-search draws a Gen Agent Trust Hub Fail, and three other sampled skills carry Snyk Warns

This is Brave's own repository: official skills for using the Brave Search API with coding agents, published under a MIT licence by the vendor that runs the endpoint. Thirteen skills, mostly one per endpoint, each a short wrapper that documents the request shape and the response you should expect back.

Two of the thirteen are where most agents should start. bx is an all-in-one CLI over the other endpoints that returns pre-extracted, token-budgeted content, and llm-context does the same job over HTTP against /res/v1/llm/context. The README marks both as recommended. The rest cover web, images, news and videos search, three local skills, suggest, spellcheck, and answers for AI-generated answers over an OpenAI-compatible chat endpoint.

It is a paid API. The README sends you to api.search.brave.com for a key, and its rate-limit note points at your plan on that dashboard; it does not describe a free tier, and this guide does not claim one.

Installation¶

The README lists your key first. Its own security note prefers agent-native config over shell profile exports, because scoped configs limit what a coding agent can read. Claude Code takes the key through settings:

{
  "env": {
    "BRAVE_SEARCH_API_KEY": "your-key"
  }
}

For any other agent, the portable version is an environment variable:

export BRAVE_SEARCH_API_KEY="your-key"

Claude Code also supports the plugin marketplace, which needs no git and auto-updates:

/plugin marketplace add brave/brave-search-skills
/plugin install brave-search-skills@brave-search

The no-git route copies every skill into the user-level directory in one line:

mkdir -p ~/.claude/skills && curl -sL https://github.com/brave/brave-search-skills/archive/main.tar.gz | tar xz -C ~/.claude/skills --strip-components=2 brave-search-skills-main/skills

There is a universal installer too, for agents outside the README's table:

npx openskills install brave/brave-search-skills

The README documents the same pattern per agent, with the skills directory swapped for .cursor/skills, .github/skills, .agents/skills (Codex), .windsurf/skills or ~/.openclaw/skills. For Hermes, copy the skills/ directories into the profile skills directory and keep the API key in the profile environment.

What It Provides¶

Twelve endpoint skills plus one CLI variant, all indexed above the table threshold.

Skill	Installs	Use For
web-search	4,714	Ranked web results with snippets and thumbnails; freshness, SafeSearch, Goggles, pagination
images-search	1,700	Image results with title, source URL and thumbnail; up to 200 per call
news-search	1,672	News articles with freshness and date-range filters, plus Goggles ranking
llm-context	1,303	Pre-extracted page text, tables and code for RAG grounding, over GET or POST
videos-search	1,252	Video results with duration, view count and creator, plus freshness filters
answers	1,245	AI-grounded answers over an OpenAI-compatible chat completions endpoint, with citations
local-pois	1,165	Full business detail for known POI IDs: ratings, hours, contact, up to 20 per call
bx	1,145	All-in-one CLI over the other endpoints, with token-budgeted output for shell work
local-descriptions	1,142	AI-written POI descriptions grounded in web context, max 20 IDs per request
spellcheck	1,134	Query correction for pre-search cleanup or a did-you-mean prompt
suggest	1,124	Query autocomplete under 100ms, typo-resilient, with optional rich entity data
local-place-search	424	Standalone place search by coordinates or location string, no POI ID needed
bx-search	12	OpenClaw packaging of the bx CLI skill under clawhub/, licence MIT-0

All 13 indexed listings appear in the table above.

Why This Matters for Hermes Agents¶

Grounding is the difference between an agent that searches and an agent that cites. llm-context returns page text already extracted for a model, so a Hermes agent can quote a source without a second fetch, parse and strip step.

Endpoint-per-skill is a clean design. The agent loads a short skill for the endpoint it needs instead of carrying a full API manual in context, which matters on a long session with a fixed window.

Goggles are the feature with no equivalent in most search APIs: a small text program that discards, boosts or downranks domains, carried by web, news, images, videos and llm-context. For a researcher agent, that is a reusable filter you write once and pass on every query.

answers speaks the OpenAI chat completions shape, so an agent already wired for an OpenAI-compatible client can point at Brave without a new SDK. Research mode adds multi-search iteration, streaming required.

Local coverage is three skills deep: local-place-search finds the place with no prior call, then local-pois and local-descriptions enrich the IDs it returns. That chain is the pattern to copy for any two-step lookup.

The caution is cost and quota. Every call is billable against a plan, and nothing in the skills caps a retry loop.

Usage¶

You say	What happens
"Search the docs for this error"	bx runs a token-budgeted context query and returns extracted page text
"What shipped in the last day?"	news-search with freshness set to the past day
"Find every page on our domain that mentions pricing"	web-search with a Goggles program that scopes and boosts the domain
"What is the best coffee near this address?"	local-place-search finds the place, then local-pois and local-descriptions enrich it
"Ground this answer in sources"	llm-context returns extracted content with URLs to cite
"Fix the typo before searching"	spellcheck returns the corrected query
"Give me an answer with citations"	answers posts to the chat completions endpoint in single-search or research mode

Verification¶

Confirm the skills landed:

ls ~/.claude/skills

Then fetch one raw file before installing anything. This path is skills/web-search/SKILL.md on branch main and returned 200 when this guide was written:

curl -fsSL https://raw.githubusercontent.com/brave/brave-search-skills/main/skills/web-search/SKILL.md | head -20

Frontmatter plus an endpoint description confirms the listing is backed by a real file. skills/llm-context/SKILL.md and clawhub/bx-search/SKILL.md are the other two paths worth checking, since the second one sits outside skills/. The install index is at https://www.skills.sh/brave/brave-search-skills.

To check the key end to end, the README's own web search call is the smallest test:

curl -s "https://api.search.brave.com/res/v1/web/search?q=agent+skills" -H "Accept: application/json" -H "X-Subscription-Token: ${BRAVE_SEARCH_API_KEY}" | head -20

Security¶

skills.sh verdicts for sampled skills (verified November 4, 2026) - re-check the security pages on skills.sh before production use:

Skill	Gen Agent Trust Hub	Socket	Snyk
web-search	Pass	Pass	Warn
llm-context	Pass	Pass	Warn
local-pois	Pass	Pass	Warn
spellcheck	Pass	Pass	Pass
bx-search	Fail	Pass	Warn

bx-search is the only Fail in the sample, and it is also the only listing that asks the agent to prefer itself over a built-in search tool. Read that skill's body before adding it to a profile, and keep the API key scoped rather than exported globally, as the README's own security note advises.

Integration with CorpusIQ¶

CorpusIQ is the connector layer that lets AI assistants read the systems a business already runs on - accounting, payments, CRM, email, files and analytics - and answer with the source named. It is read-only by design: it retrieves data and never writes to, or acts in, a connected system, which is what makes it safe to point at live accounts.

This page is part of the CorpusIQ directory, the reference we maintain for the MCP server and agent-skill ecosystem. Run alongside CorpusIQ, a third-party server like this one can be combined with a company's own business data in a single answer.

See what CorpusIQ connects

Limitations¶

Snapshot data, verified November 4, 2026: 18,032 combined installs across 13 indexed listings; 193 GitHub stars; MIT licence; last pushed 2026-10-29. Counts drift over time.
A key and a plan are required. The README documents no free tier and ties rate limits to the plan on the API dashboard, so check pricing before wiring this into a recurring loop.
bx is a wrapper, not the binary. The CLI itself ships separately from brave/brave-search-cli and is installed with the skill's own install script or an OpenClaw release download.
bx-search lives under clawhub/, not skills/, carries 12 installs against 4,714 for web-search, and is absent from the README's endpoint table. It is an OpenClaw packaging of the same CLI under MIT-0.
The README's Available Skills table documents 12 skills while the install index carries 13 listings; the extra one is bx-search.
Every request is billable, and the skills carry no caching or retry cap. An agent that re-searches on a failed parse spends quota without a signal.

Pro Tips¶

Start with llm-context for grounding and bx for shell work. The README marks both as recommended, and they cover most agent traffic.
Keep the key in agent-native config, as the README's security note says. A shell export is inherited by every process the agent spawns.
Use Goggles to cut noise before paginating. A three-line $discard and $site program is cheaper than filtering results after the fact.
Cache results inside a task. The same query rarely changes within a run, and every repeat call is billed.
Skip spellcheck mid-flow. Most search endpoints already correct queries, so it earns its place only in a pre-search or did-you-mean UI.

← Skills Catalog | Skills Marketplace → Powered by CorpusIQ