Parallel Search
Give an AI agent live web search and page extraction. Two read-only tools, no API key needed to start, and a key or OAuth for production limits — schemas read live 2026-09-08.
Verified connector
Listed by Anthropic as a partner connector in its Connectors Directory.
Connection checked by Agentman on .
Anthropic states this reflects the level of review a connector received, not a security audit.
Connect Parallel Search via MCP
https://search.parallel.ai/mcpWorks in any MCP-compatible client. In Agentman Studio it is one click — no config file to edit.
Use in Agentman
Connect Parallel Search once and your agents call these tools on their own — on a schedule, in a workflow, with nobody at the keyboard.
Open in Agentman StudioParallel Search Tools & Capabilities (2)
Purpose: Perform web searches and return LLM-friendly results, including excerpts that are usually sufficient to answer directly without a follow-up fetch. Ideal Use Cases: - Answering questions that require fresh or current information - Research, comparison, documentation, and troubleshooting questions - Broad tasks where multiple `search_queries` can be issued in a single call
Purpose: Fetch and extract relevant content from specific web URLs. Use only when web_search excerpts are insufficient for the task at hand. Ideal Use Cases: - The user asked about a specific URL or page - You need exact wording or quotes that excerpts may have truncated - You need full-page analysis (long article, document, or page structure) - web_search excerpts are conflicting or clearly insufficient to answer
Read from the server on 2026-09-08, including each tool's own safety annotations.
Limits
- It reads the web; it writes nothing. Both tools are annotated read-only and non-destructive by the server. There is no write path on this surface at all.
- Everything it returns is untrusted third-party content. Both tools declare
openWorldHint: true— the server's own statement that results come from an open-ended external source. Any text on a fetched page can reach your model, including text shaped like instructions. Treat results as data, never as directives, and keep the connector away from agents holding credentials or write access. Parallel's MCP page carries no security or prompt-injection section. - Anonymous use is deliberately constrained. Parallel describes the free tier as having lower rate limits, running in
fastmode with server-managed settings, and silently ignoring every connection-level override. A misconfigured anonymous client does not error — it just gets defaults. - The "token-efficient search" claim is Parallel's, not our measurement. It is Parallel's own description in Anthropic's directory listing and its documentation. What we can point at is the mechanism behind it: a documented cap of roughly 25,000 characters of excerpts per tool call, and an excerpt-first default on
web_fetch. We ran no token benchmark. - Rate limits are published for the APIs, not for anonymous MCP. Parallel documents 600 requests per minute each for its Search and Extract endpoints. It publishes no anonymous MCP figure, and we did not probe for one — finding a rate limit means deliberately exceeding it.
turbomode is narrower than the others. Parallel states thatturbocurrently supports English and Japanese queries only, and that domain and path filters are unsupported in it.- Commercial terms are reported by reference only. Parallel publishes per-mode Search pricing and a per-URL Extract rate on its pricing page. Pricing changes without notice and we make no claim about what any tier costs — read it from Parallel.
- Results depend on Parallel's retrieval, not the whole web. Absence from a result set is not evidence a page does not exist.
- We did not exercise either tool. Our check was a read-only handshake. Every behavioural claim above comes from the server's own tool descriptions and schemas or from Parallel's documentation.
Frequently asked questions
No, not to start. Parallel states the Search MCP is free to use with no API key required, and calls it suitable for exploration and light use. Our anonymous handshake on 2026-09-08 returned both tool schemas with no credential. Parallel directs production use to a key or OAuth, which raises the rate limits and unlocks connection-level search settings.
Authentication is the only difference. Parallel documents that the default endpoint answers anonymously and does not advertise OAuth, so an OAuth-capable client never prompts you to sign in there. The mcp-oauth endpoint requires either a Bearer key or the OAuth flow and returns 401 to anonymous requests, which is what Parallel recommends for organisation-wide or zero-data-retention deployments.
No. Both tools are annotated readOnlyHint true and destructiveHint false by the server itself, read live on 2026-09-08. There is no write path, no object model and no account state to modify. Both also declare openWorldHint true, meaning the results come from arbitrary third-party web pages rather than a bounded system you control.
Roughly 25,000 characters of excerpts, which Parallel says is capped deliberately to stay inside typical MCP client output limits. The web_fetch tool accepts up to 20 URLs per request and defaults to excerpt mode. Setting full_content to true returns whole pages as markdown, which Parallel warns can run to tens of thousands of tokens and exceed a client's output limit.
Usually the wrong file or the wrong format. Parallel's troubleshooting notes that VS Code reads a workspace .vscode/mcp.json and a profile-level file, but never a home-directory copy, and reports no error when a server sits in the wrong place. VS Code also uses a top-level servers object with a type field, not Cursor's mcpServers shape.
Parallel publishes a default quota of 600 requests per minute for its Search API and 600 per minute for Extract, the two endpoints these tools call. It does not publish a separate figure for anonymous MCP use, describing that tier only as having lower rate limits. Parallel asks teams needing higher limits to contact support with their use case.
Yes, but only on an authenticated connection. Parallel documents connection-level overrides set through URL query parameters or a configuration header, covering mode, location, result count and source policy. Parallel states these overrides are ignored for anonymous free-tier requests, and that domain or path filters need fast, basic or advanced mode rather than turbo.
Sources
- Parallel Docs, "Search MCP" (retrieved 2026-09-08, HTTP 200, served as clean Markdown by appending
.mdto the page URL). This is the exactdocumentationvalue in Anthropic's directory entry. Source for both endpoints, the free-tier posture, the 25,000-character excerpt cap, the per-call parameter tables, the connection-level override mechanism and validation behaviour, the per-client setup instructions and the troubleshooting section. · retrieved 2026-09-08 - Parallel Docs, "Search Modes" (retrieved 2026-09-08, HTTP 200). The four modes with their stated latencies, the
advanceddefault for the API, andturbo's English/Japanese language limitation. · retrieved 2026-09-08 - Parallel Docs, "API Rate Limits" (retrieved 2026-09-08, HTTP 200). The 600-per-minute quotas for Search and Extract, and the instruction to contact support for higher limits. · retrieved 2026-09-08
- Parallel Docs, "Parallel API Pricing" (retrieved 2026-09-08, HTTP 200). Consulted; cited by reference only, with no figure restated as our own claim. · retrieved 2026-09-08
- Parallel documentation machine index (retrieved 2026-09-08, HTTP 200). Used to resolve real page slugs rather than guessing them. · retrieved 2026-09-08
- Live
tools/listhandshake againsthttps://search.parallel.ai/mcp— two tools with full descriptions, input schemas, output schemas and all three safety annotations, answered anonymously with no credential and no authentication challenge (2026-09-08). Schema committed atdocs/marketing/data/tool-schemas/parallel-search.tools.json. · retrieved 2026-09-08 - Anthropic Connectors Directory entry — . Tier, category, endpoint, transport, author and both listed tool names, read from our committed directory snapshot dated 2026-09-08. The directory's
auth_postureofno_authagrees with what we observed. · retrieved 2026-09-08 - Parallel support — <support@parallel.ai> · Privacy
Server Info
- Category
- Developer Tools
- Developer
- Parallel Search
- Tools
- 2
- Domain
- search.parallel.ai
Using Claude Desktop or another MCP client? Setup docs — the connection URL above works anywhere.
Similar MCP Servers for Developer Tools
Related Resources
Ready to connect Parallel Search?
Connect Parallel Search once and your agents call these tools on their own — on a schedule, in a workflow, with nobody at the keyboard.