Grok Search

Grok Search

Node scripts for web search, URL fetch, and lightweight site mapping. Run npm install once if dependencies are not installed, then run the scripts directly — there is no MCP server or extension to call.

Install Or Configure (agent wizard)

When the user asks to install / configure / setup this skill, follow README.md →「Agent 安装与配置向导」 (English: README.en.md →「Agent setup wizard」). Do not improvise a shorter flow.

Order:

1. Self-check Node >=18.17, skill path, npm install, existing ~/.config/grok-search/config.json (mask any key).
2. Ask protocol: chat (grok2api/NewAPI; grok-4.3-fast / grok-4.20-*) vs responses (CPA-style; often grok-4.5 only).
3. Ask apiUrl (base only) and apiKey (never echo full key).
4. Ask model id; write ~/.config/grok-search/config.json with mode 600.
5. Ask whether to test. Only if yes, run:

node "$SKILL_ROOT/scripts/search.js" --no-extra "随便搜个新闻"

1. Success if exit 0, no error, diagnostics.grok_endpoint matches choice, non-empty answer.text.
2. On failure, branch by error / HTTP status / empty search (401 key, 404 protocol/URL, 422 model, quota, proxy, modelnotfound, no citations) and ask the user for the specific fix — then retest only if they agree.

If search later returns GROK_API_URL 未配置 or GROK_API_KEY 未配置, run this wizard instead of inventing env vars.

Choose The Script

Decide before running, by what the user already gave you:

• They gave a URL and asked what it says → scripts/fetch.js.
• They named a site but no URL, and want to know what is on it → scripts/map.js, then fetch.js on the URLs you pick.
• They want current/latest information, or the URL is unknown → scripts/search.js.

Do not chain map → fetch → search by default. Run the fewest commands that answer the question. If sub-questions are independent (different sites, unrelated facts), launch the commands in parallel instead of sequentially.

Commands

./scripts/search.js "query"
./scripts/search.js --search-endpoint chat --model grok-4.3-fast "query"
./scripts/search.js --platform GitHub "query"
./scripts/search.js --extra 10 "query"
./scripts/search.js --no-extra "query"
./scripts/search.js --source-chars 200 "query"
./scripts/search.js --responses-openrouter-engine exa "strict web-only query"
./scripts/search.js --responses-x-search --responses-allowed-x-handles xai,OpenAI "query"

./scripts/fetch.js https://example.com
./scripts/fetch.js --provider direct https://example.com

Use ./scripts/fetch.js --max-chars 50000 URL only for an explicit deep read after the preview shows the page is worth reading.

./scripts/map.js https://docs.example.com --limit 20
./scripts/map.js --provider direct https://docs.example.com
./scripts/map.js https://docs.example.com --instructions "only API reference pages" --max-depth 2

Search supports two Grok protocols via searchEndpoint / --search-endpoint:

• responses (default): {apiUrl}/responses + agentic web_search. Typical for CPA / xAI-style search (e.g. grok-4.5).
• chat: {apiUrl}/chat/completions + web_search tools. Typical for grok2api / NewAPI with grok-4.3-fast, grok-4.20-fast, or grok-4.20-0309-non-reasoning.

On many relays, Responses search only works for CPA grok-4.5; prefer chat for the fast Grok models above. Both paths run alongside independent Tavily and Firecrawl searches. Tavily is used when its key is configured; Firecrawl works keyless and automatically uses FIRECRAWL_API_KEY when available. The default combined extra target is 6. Add --extra 10 only for a broader candidate-source sweep. Extras are never fed into Grok.

If Grok quota is explicitly exhausted, search.js may return a visibly marked degraded answer made from raw Tavily/Firecrawl results. Check diagnostics.degraded and diagnostics.grok_error. Other Grok failures remain errors. --no-extra disables this fallback as well as the external searches.

Reading Results

Each script writes a single JSON object to stdout. On failure it still writes JSON to stdout, a short message to stderr, and exits non-zero.

Check in this order:

• error — if present, read error.message, error.code, error.preview if present, and diagnostics.provider_attempts. Before retrying, change something: a sharper query, a different --provider (fetch/map), or a different --model (search). Do not rerun the same command.
• diagnostics.warnings and diagnostics.provider_attempts — these tell you which providers were skipped, failed, or produced content.
• Search success: read answer.text, then sources.merged. Source cards are short and use snippet, not description or content. Full source/provider raw is in sources.raw_path; read it in chunks only when needed.
• Search: inspect diagnostics.grok_endpoint, diagnostics.degraded, diagnostics.cost_usd, provider attempts, and sources.grok[].source_type (citation vs searched) before treating sources as evidence.
• Fetch success: read content.text. If content.truncated is true and the preview is enough, stop. If more is needed, read content.full_path in chunks or rerun once with a deliberate larger --max-chars.
• Map success: read urls, choose the best candidates, then fetch only the few URLs you need.

Search and fetch are intentionally separated. Use search to discover and compare sources, then fetch a specific URL for deep reading. In one research turn, fetch 1-2 URLs by default; do not batch-fetch many pages unless the user explicitly asks.

Providers And Limits

fetch.js provider order for --provider auto: Tavily Extract → Firecrawl Scrape → Direct Fetch. map.js provider order for --provider auto: Tavily Map → Direct Map.

For search, Tavily requires a key while Firecrawl uses its keyless tier by default. For fetch, missing Tavily means Firecrawl Keyless runs before Direct Fetch. The direct providers do not execute JavaScript, log in, use cookies, parse PDFs, or bypass anti-bot. Direct Map only reads /sitemap.xml and same-domain homepage links, ignores --instructions, and is limited to --max-depth 1.

Proxy

The scripts automatically use terminal proxy environment variables via undici when present: HTTP_PROXY / HTTPS_PROXY / ALL_PROXY and lowercase variants. NO_PROXY is honored, and loopback hosts are bypassed. Use GROK_PROXY="http://127.0.0.1:7890" to set a proxy just for this tool, or GROK_PROXY=off to force direct connections. If proxy debugging is needed, set GROK_DEBUG=true.

If search.js returns GROK_API_URL 未配置 or GROK_API_KEY 未配置, the project itself is missing required setup — point the user at README.md instead of trying to work around it.

When To Plan First

For multi-part research, conflicting sources, or anything that needs both site discovery and page content, read references/planning.md before running commands.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: Autsunset
• Source: Autsunset/grok-search
• License: MIT

Install and usage instructions live in the source repository linked above.