> ## Documentation Index
> Fetch the complete documentation index at: https://developers.arg.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Parallel

> Search and read the web against a plain-language objective, and run deep research tasks.

Parallel (Parallel Web Systems) connects with an API key from the Parallel platform. One key covers all three surfaces: search, extraction, and asynchronous task runs.

What makes it different from a keyword search API is the **objective**: you describe what you are trying to learn, and Parallel selects the excerpts that answer it - so one call replaces several keyword searches.

## What you can do

**Actions** - search the web, read pages, start a task run, read a task run.

**Triggers** - none. Use it through the assistant in chat or as an action node in an automation.

## Get your credential

<Steps>
  <Step title="Open the Parallel platform">
    Sign in at [platform.parallel.ai](https://platform.parallel.ai).
  </Step>

  <Step title="Generate an API key">Generate a key and copy it.</Step>
</Steps>

## Connect

<Steps>
  <Step title="Open Integrations">
    In arg.ai, go to **Settings → Integrations** and click **Connect** on **Parallel**.
  </Step>

  <Step title="Paste the key">Enter your API key in the API key field.</Step>
  <Step title="Save">arg.ai checks the key without spending a search, then connects it.</Step>
</Steps>

## Use it

* **In chat:** *"Research this company's recent funding with Parallel and cite the sources."*
* **In an automation:** add a **Parallel** action node to start a task run per row of a spreadsheet, then read each run's structured output back into the sheet.

## Search modes

Pick a `mode` for the latency and cost you need: `turbo` is the fastest and cheapest, `basic` suits most agent workloads, `fast` fits a one-second budget, and `advanced` (the default) is the deepest. Narrow the sources with `include_domains` / `exclude_domains` and `after_date` only when you must - a source policy can cut result quality by excluding relevant pages.

## Task runs are asynchronous

**Start a task run** returns a run id immediately. Choose a `processor` for the depth you need - `lite` and `base` for simple enrichments, `core` for cross-referenced answers, `pro` and `ultra` for exploratory research - and append `-fast` to any of them to trade data freshness for latency. The deeper tiers run for minutes to hours.

Read it back by id with **Read a task run**. While the run is still working, it returns the run's `status`; once that is terminal, the same read returns the finished output with its citations and confidence, so a polling automation only needs the one action.