Parallel MCP 2.0 Packs Four Web Research Tools Behind One Agent Endpoint
Parallel MCP 2.0 ships four tools behind one URL: web_search, web_fetch, web_decisions, web_research Free and keyless for search, fetch, and decisions at lower rate limits web_research requires OAuth sign-in or an API key, billed at underlying API rates Decisions returns typed answers:
Takeaways
- Parallel MCP 2.0 ships four tools behind one URL: web_search, web_fetch, web_decisions, web_research
- Free and keyless for search, fetch, and decisions at lower rate limits
- web_research requires OAuth sign-in or an API key, billed at underlying API rates
- Decisions returns typed answers: probabilities, category picks, or scores with citations
- Admins can pin search mode, exclude domains, and lock tool arguments via
x-parallel-mcp-configheader - Replaces the standalone Search MCP for new installs; Task MCP still handles async deep research
Parallel MCP 2.0 is a Model Context Protocol server for search, page extraction, structured decisions, and cited research. Its expanded tool menu lets agents match each request to a backend with a distinct output format, latency, and cost profile.
One endpoint routes four jobs
The server runs at https://mcp.parallel.ai/mcp . Each tool connects to a separate Parallel API and handles a specific stage of web research.
| Tool | Backend | Behavior |
|---|---|---|
| web_search | Search API | Returns ranked excerpts that often contain the requested answer. It uses fast mode by default. |
| web_fetch | Extract API | Reads specified URLs and returns relevant sections or complete pages as Markdown. |
| web_decisions | Decisions API | Returns typed judgments, including probabilities, category choices, and scores, with supporting sources. |
| web_research | Responses API | Researches one question end to end, waits synchronously for completion, and returns a cited answer in the same call. |
Selecting a tool also selects its latency and billing profile. An agent can use a quick search for a factual lookup, fetch known pages for extraction, request a constrained judgment, or run a complete research task.
Keyless access covers three tools
web_search , web_fetch , and web_decisions run without an API key under lower free rate limits. web_research always requires authentication.
Claude Code and Codex can register the server with one command:
OAuth sign-in uses https://mcp.parallel.ai/mcp?login . Applications can also authenticate with an API key in either the Authorization: Bearer or x-api-key header.
Search pricing follows the underlying API rates:
fastandturbo: $1 per 1,000 requestsbasicandadvanced: $5 per 1,000 requests
Every tool result includes its cost in _meta["parallel/usage"] , allowing applications and gateways to attribute spending to individual calls.
Decisions turns evidence into typed output
web_decisions accepts background information in state and as many as 100 entries in questions . It searches for evidence and returns a typed answer with citations for each question.
| Question type | Input and result |
|---|---|
| noul | Evaluates a true-or-false statement and returns a probability from 0 to 1. |
| choice | Selects among 2 to 255 named categories and returns a probability for each option. |
| score | Evaluates 2 to 10 ordered criteria and returns a probability-weighted mean. |
These structured results can support account qualification, policy checks, vendor comparisons, and other workflows that need evidence converted into constrained outputs. The tool removes the application-side step of combining search results into a probability, category, or score.
Parallel classifies web_decisions as an early release, and models may not select it reliably on their own. The documentation recommends naming Parallel Decisions explicitly in prompts and stating the fact or criterion it should evaluate.
Admins can pin the model’s choices
Administrators can enforce tool arguments with a JSON object in the x-parallel-mcp-config header. Values under override are applied after the model’s arguments, preventing the model from changing them. A model_choice list identifies the fields that remain available as per-call tool arguments.
The following configuration fixes every search to basic mode, excludes reddit.com , and allows the model to add included domains for an individual query:
A tools query parameter can narrow the exposed menu. For example, ?tools=web_search,web_decisions prevents an agent from invoking fetch or billed research calls. Parallel enforces overrides on its servers, including requests routed through gateways such as LiteLLM.
Best fits and current limits
Agents that need current, cited web data can use the server without operating crawler or scraping infrastructure. Parallel identifies several common patterns:
- Personal agents: checking current prices, opening hours, and availability
- Research agents: producing briefs with source citations
- Sales agents: qualifying companies and evaluating accounts
- Coding agents: retrieving current documentation beyond model training data
Current product constraints affect language coverage, retained context, and tool selection:
turbomode currently supports queries in English and Japanese.- Organizations using Zero Data Retention cannot run follow-up research from an earlier
response_id. - Models may require explicit prompting to select the early-release
web_decisionstool.
MCP 2.0 is the recommended server for new installations and replaces the standalone Search MCP and Task MCP in that path. Both standalone servers remain operational. Parallel continues to direct asynchronous research reports and list-enrichment jobs to the Task MCP or the Parallel CLI .
Production behavior becomes policy
Server-enforced configuration lets a team define search modes, freshness windows, source rules, and tool availability once. Connected agents inherit those settings, and per-call usage metadata records their spending. The result is consistent, auditable web research behavior across an agent fleet.
Reporting is indexed from AlphaSignal. Rights remain with the original publisher and cited sources.
Read original report
