Imported from modbender/skill-library-mcp (data/parallel-ai-search/SKILL.md). Install upstream with npx skills add modbender/skill-library-mcp --skill parallel-ai-search. Copyright stays with the author.

Parallel AI Search (OpenClaw skill)

Use this skill to run web research through Parallel Search (ranked, LLM-optimised excerpts) and Parallel Extract (clean markdown from specific URLs, including JS-heavy pages and PDFs).

The skill ships tiny Node .mjs helpers so the agent can call the APIs deterministically via the OpenClaw exec tool.

Quick start

1) Provide the API key

Prefer configuring it in ~/.openclaw/openclaw.json (host runs):

{
  skills: {
    entries: {
      "parallel-ai-search": {
        enabled: true,
        apiKey: "YOUR_PARALLEL_API_KEY"
      }
    }
  }
}

Notes:

• If the agent run is sandboxed, the Docker sandbox does not inherit host env. Provide the key via agents.defaults.sandbox.docker.env (or bake it into the image).
• This skill is gated on PARALLEL_API_KEY. If it’s missing, OpenClaw won’t load the skill.

2) Run a search

Use exec to run:

node {baseDir}/scripts/parallel-search.mjs \
  --objective "When was the United Nations established? Prefer UN websites." \
  --query "Founding year UN" \
  --query "Year of founding United Nations" \
  --max-results 5 \
  --mode one-shot

3) Extract content from URLs

node {baseDir}/scripts/parallel-extract.mjs \
  --url "https://www.un.org/en/about-us/history-of-the-un" \
  --objective "When was the United Nations established?" \
  --excerpts \
  --no-full-content

4) One command: search → extract top results

node {baseDir}/scripts/parallel-search-extract.mjs \
  --objective "Find recent research on quantum error correction" \
  --query "quantum error correction 2024" \
  --query "QEC algorithms" \
  --max-results 6 \
  --top 3 \
  --excerpts

When to use

Trigger this skill when the user asks for:

• “Parallel search”, “parallel.ai search”, “Parallel Extract”, “Search API”, “Extract API”
• “web research with Parallel”, “LLM-optimised excerpts”, “source_policy/include_domains”, “after_date”, “fetch_policy”
• “extract clean markdown from URL/PDF”, “crawl a JS-heavy page”, “get fresh web results”

Default workflow

1. Search with an objective + a few search_queries.
2. Inspect titles/URLs/publish dates; choose the best sources.
3. Extract the specific pages you actually need (top N URLs).
4. Answer using the extracted excerpts/full content.

Use Search to discover; use Extract to read.

Best-practice prompting for Parallel

Objective

Write 1–3 sentences describing:

• the real task context (why you need the info)
• freshness constraints (“prefer 2025+”, “after 2024-01-01”, “use latest docs”)
• preferred sources (“official docs”, “standards bodies”, “GitHub releases”)

search_queries

Add 3–8 keyword queries that include:

• the specific terms, version numbers, error strings
• common synonyms
• if relevant, date terms (“2025”, “2026”, “Jan 2026”)

Mode

• Use mode=one-shot for single-pass questions (default).
• Use mode=agentic for multi-step research loops (shorter, more token-efficient excerpts).

Source policy

When you need tight control, set source_policy:

• include_domains: allowlist (max 10)
• exclude_domains: denylist (max 10)
• after_date: RFC3339 date (YYYY-MM-DD) to filter for freshness

Scripts

All scripts print a JSON response to stdout by default.

scripts/parallel-search.mjs

Calls POST https://api.parallel.ai/v1beta/search.

Common flags:

• --objective "..."
• --query "..." (repeatable)
• --mode one-shot|agentic
• --max-results N (1–20)
• --include-domain example.com (repeatable)
• --exclude-domain example.com (repeatable)
• --after-date YYYY-MM-DD
• --excerpt-max-chars N (per result)
• --excerpt-max-total-chars N (across results)
• --fetch-max-age-seconds N (force freshness; 0 disables)
• --request path/to/request.json (advanced: full request passthrough)
• --request-json '{"objective":"..."}' (advanced)

scripts/parallel-extract.mjs

Calls POST https://api.parallel.ai/v1beta/extract.

Common flags:

• --url "https://..." (repeatable, max 10)
• --objective "..."
• --query "..." (repeatable)
• --excerpts / --no-excerpts
• --full-content / --no-full-content
• --excerpts-max-chars N / --excerpts-max-total-chars N
• --full-max-chars N
• --fetch-max-age-seconds N (min 600 when set)
• --fetch-timeout-seconds N
• --disable-cache-fallback
• --request path/to/request.json (advanced)

scripts/parallel-search-extract.mjs

Convenience pipeline:

1. Search
2. Extract the top N URLs from the search results (single Extract call)

Common flags:

• All parallel-search.mjs flags
• --top N (1–10)
• Extraction toggles: --excerpts, --full-content, plus the extract excerpt/full settings

Output handling conventions

When turning API output into a user-facing answer:

• Prefer official / primary sources when possible.
• Quote or paraphrase only the relevant extracted text.
• Include URL + publish_date (when present) for transparency.
• If results disagree, present both and say what each source claims.

Error handling

Scripts exit with:

• 0 success
• 1 unexpected error (network, JSON parse, etc.)
• 2 invalid arguments
• 3 API error (non-2xx) — response body is printed to stderr when possible

References

Load these only when needed:

• references/parallel-api.md — compact API field/shape reference
• references/openclaw-config.md — OpenClaw config + sandbox env notes
• references/prompting.md — objective/query templates and research patterns