How to Use OpenClaw Brave Search

Brave Search is one of the structured web_search providers OpenClaw can use for grounded retrieval. It is a good default when you want normal search results with titles, URLs, and snippets, plus optional LLM Context mode for richer source chunks. The key is to treat it as a paid API-backed provider with configuration, filters, and usage limits, not as a free infinite search box.

30-second answer

Create a Brave Search API key, set BRAVE_API_KEY or plugins.entries.brave.config.webSearch.apiKey, and choose Brave under tools.web.search.provider when you want it pinned. The web_search tool then accepts query, count, country, language, search_lang, ui_lang, freshness, date_after, and date_before. Use mode web for normal results or llm-context for grounded chunks and source entries.

Where it fits

Use Brave Search for current web research, competitor scans, documentation discovery, and lightweight SERP checks. Pair it with web_fetch when you need to read a specific result. Search finds candidates; fetch extracts one page. Keeping those roles separate makes agent research easier to audit because the answer can show which source was found and which URL was actually read.

Docs-grounded facts

• Brave Search uses BRAVE_API_KEY or plugin-scoped webSearch apiKey.
• webSearch.mode supports web and llm-context.
• web_search count supports 1 to 10 results.
• Brave supports country, language, search_lang, ui_lang, freshness, date_after, and date_before in normal mode.
• llm-context mode returns grounded source entries instead of normal snippet shape.
• Results are cached for 15 minutes by default.

Set it up deliberately

The canonical provider config lives under plugins.entries.brave.config.webSearch. The legacy tools.web.search.apiKey compatibility path still loads, but it is not the preferred place for new setup. In tools.web.search, set provider to brave, tune maxResults and timeoutSeconds, and leave cacheTtlMinutes at a reasonable value unless you truly need repeated fresh searches.

Use it safely

Brave has plan costs. The docs note the Search plan cost and the monthly credit model, and they recommend setting usage limits in the Brave dashboard to avoid surprises. LLM Context mode does not support every normal filter; ui_lang, freshness, date_after, and date_before are not supported there. Also remember that search snippets are not source-of-truth facts. Fetch and verify important pages.

Common mistakes

The common mistake is using country and language casually, then wondering why results differ between runs. Set filters only when locale matters. Another mistake is using llm-context mode for a query that needs freshness filters. If time filtering is important, use normal web mode. If source chunks are more important than date filters, use llm-context and document that tradeoff.

Verification checklist

Run a small query, confirm result count and locale behavior, then fetch one result to prove the retrieval chain works. Check the tool output for URLs, not just titles. For cost control, monitor dashboard usage after automation starts, especially if a cron or agent workflow can search repeatedly.

Playbook angle

The OpenClaw Playbook treats search as a budgeted research capability. Brave is strong when you pin it deliberately, cache sanely, and combine it with web_fetch. That gives your agent grounded web context without turning every answer into an uncontrolled browsing session.

Operator note

How to Use OpenClaw Brave Search works best when it is written into a small runbook instead of left as tribal knowledge. Record the intended owner, the exact config surface, the channel where results should appear, the allowed inputs, the expected output, and the rollback step. OpenClaw gives agents broad tools, but the durable value comes from making each tool boring, repeatable, and auditable. I would rather have one well-scoped Brave search workflow that survives a restart than five clever demos nobody can safely run next week. If the runbook cannot explain when not to use it, keep refining before automation becomes default.