Parallel AI Skill

High-accuracy web search and research via Parallel.ai API. Optimized for AI agents with rich excerpts and citations.

Security Scan

The package largely matches its claimed search/research purpose, but contains inconsistencies (undeclared required credentials, a hard-coded API key, and undeclared binary dependencies) that make it suspicious and worth manual review before use.

Purpose & Capability

Name/description align with the included scripts (search, extract, task, monitor, findall). However the registry metadata declares no required env vars or binaries while almost every script requires PARALLEL_API_KEY (and some optionally BROWSERUSE_API_KEY). The shell script also uses curl/jq but the skill metadata does not list required binaries.

Instruction Scope

Runtime instructions and scripts stay within the stated goal of interacting with Parallel.ai (search, extraction, tasks, monitors). They do not attempt to read unrelated local files or secrets, but they do allow sending data to arbitrary webhook URLs (monitor.create) and can be configured to use third-party authenticated browsing (browser-use.com). Those features are plausible for a monitoring/research tool but expand the blast radius and require care when supplying webhook URLs or browser-use credentials.

Install Mechanism

No install spec is provided (instruction-only/inline usage), which is low-risk, but the package includes multiple executable scripts. Because the skill ships code but has no declared install steps, users may run the included scripts directly; the scripts call external network endpoints. No external downloads or obscure URLs are used.

Credentials

Discrepancy between metadata (no required env vars) and code: scripts require PARALLEL_API_KEY and some accept BROWSERUSE_API_KEY. Critically, scripts/search.py contains a hard-coded API key string as a default value — this is unexpected for a search client and may represent a leaked/test key or a backdoor allowing requests to run under the author's account. The hard-coded key is disproportionate and creates both functional and security concerns.

Persistence & Privilege

The skill does not request always:true, does not modify other skills, and does not request persistent platform privileges. It behaves like a normal task/utility client that can be invoked by the agent.

Scan Findings in Context

[hardcoded-api-key-in-search.py] unexpected: scripts/search.py includes a hard-coded API key literal (API_KEY default set to 'y2s_m4er5i6-5qCikOLUtmnkvOYRU24eDphq_jg1'). A legitimate client would not ship a usable secret; this is unexpected and risky (possible leaked/test key or intentional remote-account usage).

[undeclared-env-vars] unexpected: Code requires PARALLEL_API_KEY (mandatory in many scripts) and optionally BROWSERUSE_API_KEY, but the skill metadata declares no required env vars. This mismatch is surprising and should be corrected/verified.

[undeclared-binary-dependency] unexpected: scripts/parallel.sh uses curl and jq; metadata lists no required binaries. Users must ensure curl/jq are available or the shell script will fail.

What to consider before installing

Do not install or run this skill blindly. Before using it: 1) Treat the hard-coded API key as suspicious — do not rely on it; prefer to set your own PARALLEL_API_KEY and remove/replace any hard-coded defaults. 2) Verify provenance — the source and homepage are unknown. Ask the publisher for official docs or an official package. 3) Expect the scripts to call external APIs (api.parallel.ai, api.browser-use.com) and to allow sending data to webhooks you supply — never provide webhooks or browser-use credentials that accept arbitrary sensitive data. 4) Ensure required local tools (curl, jq, python dependencies) are present. 5) If you already ran these scripts with your credentials, consider rotating those API keys. 6) If you need this capability but want lower risk, request an official SDK/package from the Parallel.ai vendor or use a vetted connector rather than an unknown third-party bundle.

Parallel.ai 🔬

High-accuracy web search API built for AI agents. Outperforms Perplexity/Exa on research benchmarks.

Setup

pip install parallel-web

API key is configured. Uses Python SDK.

from parallel import Parallel
client = Parallel(api_key="YOUR_KEY")
response = client.beta.search(
    mode="one-shot",  # or "fast" for lower latency/cost, "agentic" for multi-hop
    max_results=10,
    objective="your query"
)

Modes

Mode	Use Case	Tradeoff
one-shot	Default, balanced accuracy	Best for most queries
fast ⚡	Quick lookups, cost-sensitive	Lower latency/cost, may sacrifice some accuracy
agentic	Complex multi-hop research	Higher accuracy, more expensive

Quick Usage

# Default search (one-shot mode)
{baseDir}/.venv/bin/python {baseDir}/scripts/search.py "Who is the CEO of Anthropic?" --max-results 5

# Fast mode - lower latency/cost ⚡
{baseDir}/.venv/bin/python {baseDir}/scripts/search.py "latest AI news" --mode fast

# Agentic mode - complex research
{baseDir}/.venv/bin/python {baseDir}/scripts/search.py "compare transformer architectures" --mode agentic

# JSON output
{baseDir}/.venv/bin/python {baseDir}/scripts/search.py "latest AI news" --json

Response Format

Returns structured results with:

• search_id - unique search identifier
• results[] - array of results with:
  • url - source URL
  • title - page title
  • excerpts[] - relevant text excerpts
  • publish_date - when available
• usage - API usage stats

When to Use

• Deep research requiring cross-referenced facts
• Company/person research with citations
• Fact-checking with evidence-based outputs
• Complex queries that need multi-hop reasoning
• Higher accuracy than traditional search for research tasks

API Reference

Docs: https://docs.parallel.ai Platform: https://platform.parallel.ai