Parallel Web SystemsNo MCP label; decisions tool earlyVerified 2026-10-28

Parallel MCP server free web search for Claude, no API key

One hosted MCP URL that gives Claude, ChatGPT, Cursor or Codex web search, page fetch and typed web-evidence answers for free without a key, at an unpublished free-tier rate limit.

Compare Docs

Verified factsChecked 2026-10-28

Price	Keyless calls are free and report no usage. With OAuth or an API key, calls bill at API rates: Search $1 per 1,000 requests in the fast mode the MCP uses by default, Extract $1 per 1,000 URLs, Decisions $1 per 1,000 requests.⁴ [Verified 2026-10-28]
API key / auth	None for web_search, web_fetch and web_decisions. OAuth (connect to ?login) or a Bearer or x-api-key header unlocks web_research and higher limits. Our keyless initialize, tools/list and web_search calls on 2026-10-28 returned HTTP 200.¹ [Verified 2026-10-28]
Rate limit	The keyless tier is “free at lower rate limits” with no number published. In our test a second keyless search about a minute after the first returned HTTP 429 with a “free-tier rate limit” message. Paid API defaults are 600 requests a minute for Search, Extract and Decisions.² [Verified 2026-10-28]
License / terms	Parallel’s Customer Terms (effective Oct 17, 2026), linked from the server’s x-parallel-terms header, apply. They ban using outputs to train models or to build data-reselling products, and give Parallel a license to use your inputs to improve its services.³ [Verified 2026-10-28]
Version status	No stage label on the MCP server itself. Search and Extract are GA at /v1; the docs call web_decisions an early release.¹ [Verified 2026-10-28]
Tools	Keyless: web_search, web_fetch (up to 20 URLs per call), web_decisions (1 to 100 typed questions). With auth: web_research. Our tools/list returned the three keyless tools. One older docs page lists only web_search and web_fetch for keyless use.¹ [Verified 2026-10-28]

FeaturesChecked 2026-10-28

YesNoN/ANot published

Capability	Value	Note
StatusSearch and Extract are GA at /v1; the docs call web_decisions an early release.	No MCP label; decisions tool early¹ [Verified 2026-10-28 on https://docs.parallel.ai/integrations/mcp/quickstart.md]	Search and Extract are GA at /v1; the docs call web_decisions an early release.
API key requiredNo key for 3 tools. web_search, web_fetch and web_decisions answered without auth on 2026-10-28; web_research needs OAuth or a key.	No¹ [Verified 2026-10-28 by test request to https://docs.parallel.ai/integrations/mcp/quickstart.md]	No key for 3 tools. web_search, web_fetch and web_decisions answered without auth on 2026-10-28; web_research needs OAuth or a key.
Free tierFree without a key. Keyless calls are free and report no usage; keyed calls bill at API rates (Search fast $1 per 1,000).	Yes¹ [Verified 2026-10-28 on https://docs.parallel.ai/integrations/mcp/quickstart.md]	Free without a key. Keyless calls are free and report no usage; keyed calls bill at API rates (Search fast $1 per 1,000).
Rate limitKeyless limit not published (“lower rate limits”); our second keyless search within about a minute got HTTP 429. Paid API default 600/min.	Not published² [Not published by vendor: checked 2026-10-28 on https://docs.parallel.ai/getting-started/rate-limits.md]	Keyless limit not published (“lower rate limits”); our second keyless search within about a minute got HTTP 429. Paid API default 600/min.
Official MCP serverHosted, mcp.parallel.ai/mcp. Streamable HTTP; tools/list run 2026-10-28 returned web_search, web_fetch and web_decisions.	Yes¹ [Verified 2026-10-28 by test request to https://docs.parallel.ai/integrations/mcp/quickstart.md]	Hosted, mcp.parallel.ai/mcp. Streamable HTTP; tools/list run 2026-10-28 returned web_search, web_fetch and web_decisions.
Official SDKPython + TypeScript. parallel-web on PyPI (1.3.7) and npm (1.3.6), for the underlying APIs.	Yes⁵ [Verified 2026-10-28 on https://docs.parallel.ai/llms.txt]	Python + TypeScript. parallel-web on PyPI (1.3.7) and npm (1.3.6), for the underlying APIs.
Data license / termsEffective Oct 17, 2026; no model training on outputs, no data reselling; Parallel may use inputs to improve services.	Parallel Customer Terms³ [Verified 2026-10-28 on https://parallel.ai/customer-terms]	Effective Oct 17, 2026; no model training on outputs, no data reselling; Parallel may use inputs to improve services.
Commercial use allowedYes, for your business use. Allowed for the customer’s business purposes, with the training, reselling and competitive-use bans in section 2.	Yes³ [Verified 2026-10-28 on https://parallel.ai/customer-terms]	Yes, for your business use. Allowed for the customer’s business purposes, with the training, reselling and competitive-use bans in section 2.
CORS / browser useAccess-Control-Allow-Origin: *. Returned on our MCP POST on 2026-10-28, with Mcp-Session-Id exposed.	Yes¹ [Verified 2026-10-28 by test request to https://docs.parallel.ai/integrations/mcp/quickstart.md]	Access-Control-Allow-Origin: *. Returned on our MCP POST on 2026-10-28, with Mcp-Session-Id exposed.
Freshness / coverageweb_fetch fetches live by default (allow_live_fetch true); cached content can be older.	Live web; cached pages allowed¹ [Verified 2026-10-28 on https://docs.parallel.ai/integrations/mcp/quickstart.md]	web_fetch fetches live by default (allow_live_fetch true); cached content can be older.

What it does

Parallel Web Systems sells search and research APIs built for AI agents, and its hosted MCP server at https://mcp.parallel.ai/mcp can be used with no account at all. Connected without credentials, the server lists three tools: web_search, which takes a plain-language objective plus a few keyword queries and returns ranked URLs with excerpts; web_fetch, which reads up to 20 URLs per call and returns the relevant passages or the whole page as Markdown; and web_decisions, an early-release tool that answers typed questions (a probability, a choice between options, or a score) from current web evidence and cites its sources. Signing in with OAuth or sending an API key adds web_research, raises the limits and bills calls to your Parallel account at the normal API rates. This server replaces Parallel’s older standalone Search MCP and Task MCP for new installs; both of those still run.

Working exampleRun 2026-10-28

ShellRequest

# 1) start a session (keyless)
curl -sS -D - https://mcp.parallel.ai/mcp \
  -H "Content-Type: application/json" \
  -H "Accept: application/json, text/event-stream" \
  -d '{"jsonrpc":"2.0","id":1,"method":"initialize","params":{"protocolVersion":"2025-06-18","capabilities":{},"clientInfo":{"name":"test","version":"1.0"}}}'

# 2) search, reusing the mcp-session-id header from step 1
curl -sS https://mcp.parallel.ai/mcp \
  -H "Content-Type: application/json" \
  -H "Accept: application/json, text/event-stream" \
  -H "mcp-session-id: $SID" \
  -d '{"jsonrpc":"2.0","id":3,"method":"tools/call","params":{"name":"web_search","arguments":{"objective":"Find the official NPPES NPI Registry API documentation page","search_queries":["NPPES NPI Registry API","NPI registry API version 2.1"]}}}'

Response (trimmed)

HTTP responseResponse

HTTP 200, text/event-stream (run 2026-10-28, no key)
event: message
data: {"jsonrpc":"2.0","id":3,"result":{"content":[{"text":"{
  \"search_id\": \"search_0cd9e05c...\",
  \"results\": [
    {\"url\": \"https://npiregistry.cms.hhs.gov/demo-api\",
     \"title\": \"NPPES NPI Registry\",
     \"excerpts\": [\"### NPPES Read API Version 2.1 Help ... |limit | |Limit results, default = 10, max = 200 | ...\"]},
    {\"url\": \"https://npiregistry.cms.hhs.gov/api-page\", ...}, ...]}"}]}}

# A second keyless search about a minute later:
HTTP 429 {"error":{"code":-32000,"message":"You've hit the free-tier rate limit for Parallel MCP. ..."}}

Agent setup

Prompt to give your agent

Prompt

Use the Parallel MCP server (https://mcp.parallel.ai/mcp, no key). Call web_search once with a one-sentence objective and two or three short keyword queries about my question, then call web_fetch on the two most relevant URLs with the same queries. Answer in five bullet points, each with its source URL. If you get a rate-limit error, stop and tell me instead of retrying in a loop. Question: what changed in the latest release of the library I name next?

Agent / MCP-style tool snippet

Shell

# Claude Code (keyless)
claude mcp add --transport http parallel https://mcp.parallel.ai/mcp
# Codex
codex mcp add parallel --url https://mcp.parallel.ai/mcp

JSON

# Cursor: .cursor/mcp.json
{
  "mcpServers": {
    "parallel": { "url": "https://mcp.parallel.ai/mcp" }
  }
}
# Only the tools you want: https://mcp.parallel.ai/mcp?tools=web_search,web_fetch
# Higher limits + web_research: 'https://mcp.parallel.ai/mcp?login' (OAuth) or an Authorization: Bearer header

Fit and safety

Good for

One-line web search and page reading for Claude Code, Cursor or Codex with no key
Quick fact checks with typed answers and citations via web_decisions
Trying Parallel’s search quality before paying for the API

Watch out

Do not build a production feature on the keyless tier. The free limit is unpublished and, in our test, a second search within about a minute was already refused with HTTP 429.
Do not use it to train a model or to fill a database you sell. Parallel’s Customer Terms ban both uses of its outputs.
Do not send confidential questions on the keyless tier if that matters to you. The terms let Parallel use your inputs to improve its services, and the keyless tools also ask the model to send its name and a session id for analytics.
Do not treat web_decisions answers as final. Parallel calls the tool an early release and any question can come back as a refusal; check the cited sources.

Alternatives

Exa hosted MCP	Another keyless hosted search-and-fetch MCP, also with unpublished free limits.
You.com MCP (free)	Keyless hosted search MCP with a published 100-requests-a-day allowance.
TinyFish Search & Fetch	Published free daily allowances (12,000 searches, 1,000 fetches) in exchange for a free account.

FAQ

Is the Parallel MCP server free?

Yes, without a key. web_search, web_fetch and web_decisions are free at lower, unpublished rate limits; signing in or adding a key bills calls at the normal API rates.

How do I add Parallel search to Claude Code?

Run claude mcp add --transport http parallel https://mcp.parallel.ai/mcp. No key is needed for search and fetch.

What is the difference between the keyless and signed-in server?

Signed-in or keyed connections add the web_research tool, get higher rate limits and can send override settings; keyless connections cannot.

Compare

Put Parallel MCP server (free, no key) side by side with other free agent tools on keys, limits, MCP support, licensing and CORS. Open the comparison grid →

SourcesFootnotes

1docs.parallel.ai/integrations/mcp/quickstart.mdVerified 2026-10-28
2docs.parallel.ai/getting-started/rate-limits.mdVerified 2026-10-28
3parallel.ai/customer-termsVerified 2026-10-28
4docs.parallel.ai/getting-started/pricing.mdVerified 2026-10-28
5docs.parallel.ai/llms.txtVerified 2026-10-28