DocumentationWhat is MCP for AI research?
Public guide
What is MCP for AI research?
MCP lets an AI client call approved Rivul Atlas research tools for scholarly search and citation metadata through one controlled connection.
On this page
MCP in plain language
Model Context Protocol, or MCP, is an open standard for connecting an AI application to external tools and data. The AI client discovers what a server offers, reads each tool's input schema, and can call an appropriate tool during a conversation.
MCP is the connection layer. It does not make a language model more accurate by itself, and it does not replace source evaluation. The quality of a research answer still depends on the records returned, the papers opened, and the researcher's judgment.
How Rivul Atlas uses MCP
Rivul Atlas exposes a remote MCP server at https://rivul.ai/mcp. An approved AI client connects to that endpoint, authenticates to one Rivul AI workspace, discovers the available tools, and sends structured tool calls when the user asks for scholarly research help.
Atlas currently provides two read-only tools. search_papers searches normalized scholarly metadata and source locations. usage_status reports the selected workspace's Atlas MCP allowance without consuming a search credit.
The same governed scholarly search capability also has a REST interface. MCP is useful when the research task begins inside an AI client; REST is useful when a developer is building a conventional application or service integration.
A research question through MCP
Suppose a researcher asks an AI client: Find three highly cited papers about retrieval-augmented generation and return their DOI or source links.
The client identifies search_papers as the relevant Rivul Atlas tool.
It sends a bounded query and result limit to Atlas instead of inventing a bibliography from model memory.
Atlas returns structured paper records with available identifiers, citation metadata, and source locations.
The client summarizes the records for the researcher, who opens the original papers before using their findings as evidence.
This flow improves provenance because the search result and its identifiers remain visible. It does not prove that a paper supports a later claim until the relevant source text is read and evaluated.
Access and security
Rivul Atlas supports browser authorization for compatible remote MCP clients and dedicated workspace MCP keys for clients that need manual credentials. Browser authorization lets the user choose one workspace and approve bounded scopes without giving the client a Rivul AI password.
Create a separate key for each client, keep it out of source code and screenshots, and revoke it when the connection is no longer required. OAuth grants and dedicated keys are revocable from the Rivul Atlas console.
Remote MCP authorization relies on discovery metadata, audience-bound access tokens, exact redirect validation, and PKCE where the client supports the browser flow. Client behavior still varies, so use a client-specific guide rather than assuming every MCP host completes the same authorization flow.
What Rivul Atlas MCP can and cannot do
Atlas MCP can search scholarly records, return citation metadata and source locations, and report the approved workspace's usage state. Its current tools are read-only.
Atlas MCP does not read a paper automatically, guarantee that a result is relevant, validate a scientific claim, edit a Rivul AI manuscript, or provide unrestricted access to publisher full text. A metadata record is a discovery lead, not evidence for a finding.
Atlas is in a controlled beta for verified Pro and Ultimate accounts. Available clients, authentication paths, and allowances can change as interoperability testing continues.
Connect an AI client
Choose the guide that matches the client you use. Each guide records the verified setup state, the expected tool names, a bounded test query, and the disconnect procedure.
For a client without a verified guide, start from its official remote HTTP MCP documentation, use https://rivul.ai/mcp as the server URL, and confirm tool discovery before running a research search. Do not treat an OAuth consent screen alone as proof that the client can call the tools.
Something here wrong or missing? Email support@rivul.ai.