bravenews
A Claude Code skill that turns Brave's News Search API into an
AI-powered news briefing tool.
Two modes:
- Quick scan (~30s) — headlines, story clustering, source
quality filtering
- Deep briefing (~60-90s) — fetches top articles, richer
analysis per story thread
Prerequisites
- Claude Code
- A Brave Search API key
(free tier: 2,000 queries/month at https://brave.com/search/api/)
Setup
1. Get a Brave API key
Sign up at https://brave.com/search/api/ and grab your API key.
2. Add the Brave Search MCP server
In your ~/.claude/.mcp.json, add:```json
{
"mcpServers": {
"brave": {
"command": "npx",
"args": ["-y", "@brave/brave-search-mcp-server"],
"env": {
"BRAVE_API_KEY": "YOUR_KEY_HERE"
}
}
}
}
Restart Claude Code after adding this.
### 3. Install the plugin```
/install-plugin rowbradley/bravenews
Or install via the searchwave marketplace:```
/install-plugin bravenews@searchwave-marketplace
## Usage```
/bravenews AI agents 2026
/bravenews Ukraine conflict
/bravenews deep Apple earnings
Quick mode (default) returns in ~30s. Add "deep" for full article
fetching (~60-90s).
How it works```
/bravenews AI agents 2026
│
▼
┌─────────────────────┐
│ 2 parallel queries │ "AI agents 2026"
│ (entity inference) │ "AI agents enterprise adoption"
└─────────┬───────────┘
▼
┌─────────────────────┐
│ Deduplicate + filter│ Remove duplicate URLs
│ │ Blocklist: PR wires, SEO farms
│ │ Tier-weight: Reuters > TechCrunch > blogs
└─────────┬───────────┘
▼
┌─────────────────────┐
│ Story clustering │ Group articles by underlying event
│ │ One entry per story, not per article
└─────────┬───────────┘
▼
┌─────────────────────┐
│ Synthesis │ Bloomberg-style brief
│ │ Per-story sources
└─────────────────────┘
## Brave API features used
| Feature | Parameter | Why |
|---------|-----------|-----|
| News endpoint | `brave_news_search` | Structured metadata (publication, age) vs general web search |
| Freshness | `freshness: "pw"` | Precise time filtering: past day, week, month, or custom date ranges |
| Extra snippets | `extra_snippets: true` | Up to 5 additional excerpts per result — richer synthesis without fetching |
| Count | `count: 15` | 15 results per query x 2 queries = 30 candidates after dedup |
| Country | `country: "US"` | Optional — localized results for region-specific topics |
## Design decisions
Things we learned building this:
- **2 queries, not 3.** Three queries produced ~40% duplicate
results. Two queries with inferred entity/sector differentiation
gives better coverage with less overlap.
- **Source quality tiering.** Two layers: a blocklist for press
release wires and SEO farms, plus tier-weighted synthesis
(Reuters > TechCrunch > random blogs).
- **Story clustering even in Quick mode.** One entry per story,
not one entry per article. Makes the output scannable.
- **Sequential WebFetch in Deep mode.** Parallel WebFetch has a
cascade failure mode in Claude Code — if one call errors, it
kills sibling calls. Sequential is slower but reliable.
- **Per-story sources.** Links directly under each story for
in-context reading, not collected at the bottom.
- **`extra_snippets: true`** is the single most impactful API
parameter for LLM synthesis — 5x more text per result means
Quick mode can produce rich output without fetching articles.
## API observations
Honest notes from building on the Brave News Search API:
- `extra_snippets` dramatically improves snippet-only synthesis
quality — this parameter alone makes Quick mode viable as a
standalone briefing tool
- `freshness` filter is precise and reliable across pd/pw/pm
and custom date ranges
- News endpoint returns structured metadata (article age,
publication name) that general web search doesn't provide —
this is what makes source-quality filtering possible
- No server-side domain filtering — would be useful for
excluding press release wires at the API level rather than
in post-processing
- A per-result relevance/confidence score would help automated
story clustering (currently inferred from source count and
recency)
- MCP server (`@brave/brave-search-mcp-server`) sets up cleanly —
single npm package, env-based API key, no additional config
## Example output (Quick mode)```
/bravenews AI agents 2026
> Summary — McKinsey has deployed tens of thousands of
> internal AI agents, per Business Insider, while Render raised
> $100M at a $1.5B valuation to build cloud infrastructure
> specifically for AI agent deployment. The Atlantic reports
> agentic tools like Claude Code are "taking America by storm,"
> collapsing months of work into hours.
>
> Consulting firms race to measure AI agent value
> (4 sources, 1 day ago)
> McKinsey CEO Bob Sternfels disclosed tens of thousands of
> internal agents; BCG tracks value by time reclaimed and
> reinvested in higher-value work.
> - https://businessinsider.com/mckinsey-bcg-pwc-ey-ai-agents-adoption-value-consulting-industry-2026-2
>
> Render raises $100M for AI agent cloud infrastructure
> (3 sources, 1 day ago)
> Series C extension values the company at $1.5B. Building
> integrated platform to replace "fragmented mess of vendors"
> for agent deployment.
> - https://pymnts.com/news/investment-tracker/2026/render-raises-100-million-to-build-cloud-tech-for-ai-agents/
(See examples/ for full outputs)
Quota note
Each Quick invocation uses 2 API calls. Deep mode uses 2 + up to
5 WebFetch calls. On Brave's free tier (2,000 queries/month),
that's ~1,000 Quick scans/month.
License
MIT
