Skip to main content
For AI agents: a documentation index is available at https://docs.parallel.ai/llms.txt. The full text of all docs is at https://docs.parallel.ai/llms-full.txt. You may also fetch any page as Markdown by appending .md to its URL or sending Accept: text/markdown.For Search and Extract, use /v1/search and /v1/extract for all new integrations and agent tool calls. Use /v1beta/search and /v1beta/extract only when maintaining an existing integration that already uses them. The request bodies changed too: on /v1/extract, full_content, fetch_policy and excerpt_settings go inside advanced_settings, and sending the v1beta top-level excerpts, full_content and fetch_policy to /v1/extract returns a 422. On /v1/search, search_queries is required (at least one short keyword query), and max_results, source_policy and fetch_policy go inside advanced_settings; sending only objective, or sending those fields at the top level, returns a 422. Do not substitute /v1 for the documented FindAll or Ingest endpoint paths.If you are evaluating Parallel Search or comparing it with other search APIs, read How to Eval Parallel Search before you build the harness, especially its Common pitfalls section. The most common mistake is sending the raw question as both objective and the only entry in search_queries. Write a goal-style objective and 1-3 short keyword queries instead.
This page covers the standalone Task MCP at task-mcp.parallel.ai, which keeps running for existing users. For new installs, use the Parallel MCP server. The Parallel MCP server’s web_research tool answers research questions in one call, with no polling. For enriching lists of items, use the Task API or the Parallel CLI.
The Task MCP Server provides two core capabilities: deep research tasks that generate comprehensive reports, and enrichment tasks that transform existing datasets with web intelligence. Built on the same infrastructure that powers our Task API, it delivers the highest quality at every price point while eliminating complex integration work. The Task MCP comprises four tools:
  • createDeepResearch — Initiates a deep research task. Use for complex questions that need thorough multi-step research with citations, where depth matters more than latency.
  • createTaskGroup — Initiates a task group to enrich multiple items in parallel. Use for lists or tabular data (CSV uploads, company lists, entity tables) that need consistent web-sourced fields per row.
  • getStatus — Lightweight status check for an in-flight task (~50 tokens). Use this when polling instead of fetching the full result.
  • getResultMarkdown — Retrieves the final output from a deep research or task group in an LLM-friendly markdown format. Call once the task is complete.
The Task MCP Server uses an async architecture that lets agents start research tasks and continue executing other work without blocking. This allows spawning any number of tasks in parallel while continuing the conversation. Due to current MCP and LLM client limitations, you need to request the result tool call in an additional turn after results are ready. The Task MCP works best when following this process:
  1. Choose a data source - See Enrichment data sources and destinations.
  2. Initiate your tasks - After you have your initial data, the MCP can initiate deep research or task groups. See Use cases for inspiration.
  3. Analyze the results - The LLM provides a link to view progress as results come in. After completion, prompt the LLM to analyze the results and answer your questions.

One-Click Install

Install in any of the following clients with a single click.

Install in Cursor

One-click install for Cursor.

Install in VS Code

One-click install for VS Code.

Install in LM Studio

One-click install for LM Studio.

Install in Goose

One-click install for Goose.
The Task MCP requires an API key. After one-click install, add your key from platform.parallel.ai as a Bearer token in the Authorization header (configured in the client’s MCP settings after install). For Claude Code, Codex CLI, Claude Desktop, Windsurf, Zed, Gemini CLI, Warp, Kiro, and other clients, see the Installation section below — those clients use a CLI command or a JSON config file rather than deep-link URLs.

Enrichment data sources and destinations

The task group tool can be used directly from LLM memory, but is often combined with a data source. The following data sources work well with the Task Group tool:
  • Upload tabular files - Use the Task MCP with Excel sheets or CSVs. Some LLM clients (such as ChatGPT) allow uploading Excel or CSV files. Availability varies by client.
  • Connect with databases - Several MCPs allow your LLM to retrieve data from your database, such as Supabase MCP and Neon MCP.
  • Connect with documents - Documents may contain vital initial information to start a task group. See Notion MCP and Linear MCP.
  • Connect with web search data - Parallel Search MCP or other web tools can provide an initial list of items, which is often a great starting point for a task group.

Use cases

The Task MCP serves two main purposes. First, it makes Parallel APIs accessible to anyone requiring reliable research or enrichment without coding skills. Second, it’s a great tool for developers to experiment with different use cases and see output quality before writing code. Below are examples of using the Task MCP (sometimes in combination with other MCPs): Day-to-day data enrichment and research: Assisting development with Parallel APIs: This is just the tip of the iceberg. Share your most compelling use cases with us to grow this list and inspire others!

Installation

The Task MCP can be installed in any MCP client. The server URL is: https://task-mcp.parallel.ai/mcp The Task MCP can also be used programmatically by providing your Parallel API key in the Authorization header as a Bearer token.

Cursor

Add to ~/.cursor/mcp.json or .cursor/mcp.json (project-specific):
For more details, see the Cursor MCP documentation.

VS Code

The quickest path is the guided flow — no file editing required. Run MCP: Add Server from the Command Palette (⇧⌘P / Ctrl+Shift+P), pick HTTP, and paste the server URL:
VS Code then prompts for a server name (Parallel Task MCP) and a configuration destination. For the Copilot harness, choose Global → Copilot Global for a personal configuration. This writes a portable configuration with a top-level mcpServers object to $COPILOT_HOME/mcp-config.json (or ~/.copilot/mcp-config.json when COPILOT_HOME is unset). Workspace-root .mcp.json is another portable destination; the guided flow offers it when the experimental chat.mcp.workspaceRootConfig.enabled setting is enabled. Review any trust and authentication prompts before accepting them, then complete the browser sign-in when VS Code opens it — the Task MCP requires authentication and supports OAuth, so VS Code obtains and stores the token for you. The examples below use the Local harness configuration format: .vscode/mcp.json in your workspace, or the user-profile mcp.json that MCP: Open User Configuration opens. These compatibility destinations use a top-level servers object and are labeled deprecated in the guided flow. Do not paste this format into a portable configuration.
If you would rather use an API key than OAuth, do not hardcode secrets in mcp.json — the file is often committed to version control. Instead, use one of the secure approaches below.
Set PARALLEL_API_KEY in the environment that VS Code inherits—for example, export it in your shell profile (.bashrc, .zshrc, etc.) before launching VS Code:
Then reference the variable in mcp.json:
VS Code must be restarted after changing the environment variable for the new value to take effect. In a portable Copilot configuration, use ${PARALLEL_API_KEY} instead of ${env:PARALLEL_API_KEY}.
Get your API key from platform.parallel.ai.
The user-level mcp.json lives inside your VS Code profile folder, not at ~/.vscode/mcp.json. Open it with MCP: Open User Configuration rather than typing the path — a file you create at ~/.vscode/mcp.json is ignored, with no error to tell you why.
For more details, see the VS Code MCP documentation and configuration reference.

Claude Desktop / Claude.ai

Go to Settings → Connectors → Add Custom Connector, and fill in:
If you are part of an organization, you may not have access to custom connectors. Contact your organization administrator for assistance. For a local Claude Desktop configuration instead, go to Settings → Developer → Edit Config and use the following JSON after retrieving your API key from Platform. If your config already has an mcpServers object, add the Parallel Task MCP entry inside it without replacing your other servers. This file-based option does not apply to Claude.ai or bypass organization policies.
For more details, see the Claude remote MCP documentation.

Claude Code

Run this command in your terminal:
In Claude code, use the command:
Then follow the steps in your browser to login. For more details, see the Claude Code MCP documentation.

Codex CLI

The Task MCP always requires authentication. Choose one of:
API key via environment variable
OAuth
Each command writes the equivalent [mcp_servers.parallel-task] entry to ~/.codex/config.toml — you can also hand-edit it if you prefer. Restart Codex after adding the server. For more details, see the Codex MCP documentation.

Other Clients

Add to ~/.codeium/windsurf/mcp_config.json:
For more details, see the Windsurf MCP documentation.
Go to MCP Servers → Remote Servers → Edit Configuration:
For more details, see the Cline MCP documentation.
Add to ~/.gemini/settings.json:
For more details, see the Gemini CLI MCP documentation.
Custom MCP server connections are subject to account and workspace permissions. You don’t need to turn on Developer Mode first.Go to Plugins, select the plus button, then select Add custom MCP server.Enter a name, set the public HTTPS endpoint URL under Connection, and configure authentication:
Review the risk warning, select I understand and want to continue, then select Create as a plugin and review the discovered tools and metadata.Install the resulting plugin, then start a new conversation, type @ in the prompt box, and select the plugin.For more details, see the ChatGPT custom MCP server documentation.
Run this command in your terminal:
The OAuth flow will start when you start Amp.For more details, see the Amp MCP documentation.
Add to .kiro/settings/mcp.json (workspace) or ~/.kiro/settings/mcp.json (global):
For more details, see the Kiro MCP documentation.
In the Antigravity Agent pane, click the menu (⋮) → MCP Servers → Manage MCP Servers → View raw config, then add:
For more details, see the Antigravity MCP documentation.
Add to opencode.json (project) or ~/.config/opencode/opencode.json (global):
For more details, see the OpenCode MCP documentation.
Add to .roo/mcp.json in your workspace (or edit the global mcp_settings.json from the Roo Code MCP settings view):
Project config takes precedence over global. For more details, see the Roo Code MCP documentation.
The OpenHands CLI is no longer actively maintained; these commands are for existing CLI users. For a maintained local installation, OpenHands recommends Agent Canvas.Run this command in your terminal:
Inside an OpenHands conversation, use /mcp to verify the server is active. For more details, see the OpenHands MCP documentation.
Run this command in your terminal:
Inside droid, type /mcp to open the interactive manager and verify the server is connected. For more details, see the Factory MCP documentation.
Pi 0.99+ has built-in MCP support. This setup uses the pi-mcp-adapter package; see its README for a comparison with built-in MCP. Install it first:
Restart Pi, then add the Task MCP to .mcp.json in your project (or ~/.config/mcp/mcp.json for a user-global config). The Task MCP requires a Parallel API key:
Set PARALLEL_API_KEY in your environment with your key from platform.parallel.ai. directTools: true registers createDeepResearch, createTaskGroup, getStatus, and getResultMarkdown alongside Pi’s built-in tools instead of hiding them behind the adapter’s mcp proxy. Run /mcp inside Pi to verify the server is detected. For more details, see the pi-mcp-adapter README.Alternatively, skip MCP entirely and use the Parallel CLI as a Pi skill — that’s the same pattern as our ClawHub / OpenClaw integration.
OpenClaw stores MCP server definitions in ~/.openclaw/openclaw.json under mcp.servers. The Task MCP requires a Parallel API key — OpenClaw’s remote MCP schema accepts it via the headers map, so avoid committing a literal key and pass it through an env-var or secret reference.Save via the openclaw mcp set CLI:
Or edit ~/.openclaw/openclaw.json directly:
The transport field is required — OpenClaw defaults to SSE when it’s omitted, but the Task MCP uses Streamable HTTP. Replace YOUR-PARALLEL-API-KEY with a key from platform.parallel.ai. Sensitive values in url and headers are redacted in OpenClaw logs.openclaw mcp set only writes to config — it doesn’t connect to the server or reload running agents. Start a new OpenClaw agent session (or restart your current one) for the tools to show up. Verify the saved definition with openclaw mcp list / openclaw mcp show parallel-task. For the full MCP CLI reference, see the OpenClaw MCP docs.
Add to ~/.hermes/config.yaml:
For more details, see the Hermes Agent MCP documentation.
Add a file at .continue/mcpServers/parallel-task.yaml in your workspace:
For more details, see the Continue.dev MCP documentation.
Open Settings → AI → MCP Servers, click Add Server, and choose Add Remote Server. Or add the entry to your Zed settings.json (open it with the zed: open settings file action). Zed uses a top-level context_servers object:
Replace YOUR-PARALLEL-API-KEY with a key from platform.parallel.ai. To sign in with OAuth instead, leave out the headers object. Zed then prompts you to authenticate in your browser.For more details, see the Zed MCP documentation.
Open Settings → Agents → MCP servers, click + Add, and paste:
Replace YOUR-PARALLEL-API-KEY with a key from platform.parallel.ai. To sign in with OAuth instead, leave out the headers object. Warp opens a browser sign-in flow when the server starts.
Do not put a real API key in .warp/.mcp.json at the project root. That file lives with the repository and is easy to commit. For a file-based setup, use OAuth (omit headers), or keep the key in Warp’s settings UI or the global ~/.warp/.mcp.json.
For more details, see the Warp MCP documentation.
Run the Install MCP Server command and fill in:
Add an HTTP header Authorization: Bearer YOUR-PARALLEL-API-KEY with a key from platform.parallel.ai. To sign in with OAuth instead, leave the header out and keep the OAuth type on Dynamic. Raycast then prompts you to Sign In before the tools become available.For more details, see the Raycast MCP documentation.
If your client isn’t listed, check its documentation before copying a configuration. Many clients accept the mcpServers format shown in the Cursor section, but others use a different wrapper or field names (e.g., VS Code’s Local harness uses top-level servers, and Windsurf uses serverUrl).

Best Practices

Choose enabled MCPs carefully

Be careful which tools and features you have enabled in your MCP client. When using Parallel in combination with many other tools, the increased context window may cause degraded output quality. Additionally, the LLM may prefer standard web search or deep research over Parallel if both are enabled. Turn off other web or deep-research tools, or explicitly mention that you want to use Parallel MCPs.

Limit data source context size

The Task MCP is a powerful tool for batch deep research, but it is constrained by the context window size and max output tokens of the LLM. Design your prompts and tool calls to avoid overflowing these limitations, or you may experience failures, degraded performance, or lower output quality. For large datasets, use the API or other no-code integrations. The Task MCP is designed for smaller parallel tasks and experimentation.

Follow up on tasks

The Task MCP only initiates Deep Research and Task Groups—it does not wait for tasks to complete. Fetch the status or results using a follow-up tool call after research is complete. The asynchronous nature allows initiating several deep researches and task groups without overflowing the context window. To perform multiple tasks or batches in a workflow, reply each time to verify the task is complete and initiate the next step.

Use with larger models only

While our Web Search MCP works well with smaller models (such as GPT OSS 20B), the Task MCP requires strong reasoning capability. Use it with larger models only (such as GPT-6 Astra or Claude Sonnet 5). Smaller models may result in degraded output quality.

Troubleshooting

Common Installation Issues

This error occurs when Cline attempts OAuth authentication but the redirect URI isn’t using HTTPS.Solution: Use the API key approach instead of OAuth:
  1. Get your API key from platform.parallel.ai
  2. Configure Cline with the bearer token method:
Replace YOUR-PARALLEL-API-KEY with your actual API key.
Gemini CLI uses HTTP MCPs and authenticates via OAuth. If OAuth isn’t working, you can provide your API key directly.Solution: Use environment variables or the mcp-remote proxy:
Add this to ~/.gemini/settings.json and replace YOUR-PARALLEL-API-KEY with your key from platform.parallel.ai.
VS Code’s Local harness configuration uses .vscode/mcp.json inside the workspace or a user-level mcp.json in your VS Code profile folder, which MCP: Open User Configuration opens. ~/.vscode/mcp.json is not a supported configuration location. Copilot’s portable configurations instead use workspace-root .mcp.json or $COPILOT_HOME/mcp-config.json (~/.copilot/mcp-config.json by default).If a hand-written config isn’t picked up, run MCP: Open User Configuration and check whether the file that opens is the one you edited. Run MCP: List Servers to confirm VS Code sees the server, and use MCP: Add Server to let VS Code write the entry to the correct file.
VS Code supports two configuration formats. Match the top-level object to the destination file.Portable Copilot format (workspace-root .mcp.json or Copilot user mcp-config.json):
Local harness format (.vscode/mcp.json or user-profile mcp.json):
The Local harness format uses servers; the portable format uses mcpServers. Both support type: "http" for remote HTTP servers, but portable configurations do not support VS Code’s ${input:...} variables.
Windsurf uses a different configuration format than Cursor.Correct Windsurf configuration:
Note: Windsurf uses serverUrl instead of url. Add this to your Windsurf MCP configuration file.
If you’re getting connection errors:
  1. Check your network: Ensure you can reach https://task-mcp.parallel.ai
  2. Verify API key: Make sure your key is valid at platform.parallel.ai
  3. Check balance: A 402 error means insufficient credits—add funds to your account
  4. Restart your IDE: Some clients cache MCP connections
If the MCP installs but tools don’t show up:
  1. Restart your IDE completely (not just reload)
  2. Check configuration syntax: Ensure valid JSON with no trailing commas
  3. Verify the server URL: Must be exactly https://task-mcp.parallel.ai/mcp
  4. Check IDE logs: Look for MCP-related errors in your IDE’s output/debug panel