Connect with MCP

Use AI Baseline from an application that supports remote Model Context Protocol (MCP) servers. Your application retrieves evidence from a selected knowledge base through MCP and uses that evidence to answer your questions. This guide uses the S&P 500 knowledge base as its example. For steps specific to your application, use Connect with Claude or Connect with ChatGPT.

Select a knowledge base

The MCP server URL has the shape https://mcp.ai-baseline.dev/{org}/{base}. It identifies one knowledge base and remains unversioned. The REST API uses https://api.ai-baseline.dev/{org}/{base}/v2; copy the MCP URL when configuring an MCP client. Use the exact organization and knowledge-base identifiers provided for your connection. For another base, replace the example URL and connection name below, and use a key that grants access to that base. Each base needs its own connection and authorization, even when several bases use the same domain. See organizations, knowledge bases, and domains.

Connection settings

Setting	Value
Suggested name	AI Baseline S&P 500
Server URL	`https://mcp.ai-baseline.dev/baseline/sandp500`
Transport	Streamable HTTP over HTTPS
Authentication	OAuth authorization code flow with PKCE S256
Client registration	CIMD or DCR
Evidence permission	`evidence:read`

The server URL is the complete MCP endpoint. Copy it exactly into your client’s remote server settings. The /connect address is a setup page; use the URL in the table for the connection itself.

Add and authorize the connection

Open your application’s settings for remote MCP servers and add a connection with the name and server URL above.
Choose Streamable HTTP if the application asks for a transport, and OAuth for authentication. Let the client discover the authorization server and use a supported client registration method described below.
Complete the browser authorization flow. Enter your AI Baseline API key on the AI Baseline sign-in page, then review and approve access to the collection. Automatic client setup leaves optional client ID and secret fields empty; your API key is the credential for the sign-in page.
Enable the connection’s tools in your application. Ask it to use AI Baseline S&P 500 to describe the available filings and date ranges. Confirm that an authenticated tool call succeeds.
Ask your question, for example: What supply chain risks did Apple disclose in its 10-K filings? Cite the supporting evidence and identify the filing years.

Your application needs to support both the remote transport and browser-based OAuth authorization. If it only supports local process servers or fixed API-key headers, use a client with these capabilities or the HTTP API quickstart.

Client registration

AI Baseline supports both Client ID Metadata Documents (CIMD) and Dynamic Client Registration (DCR). These mechanisms identify the application connecting to AI Baseline. Your API key and consent authorize that application to access the collection.

Method	How it works	When to use it
CIMD	The client supplies an HTTPS URL as its client ID. AI Baseline reads the published document to validate the client’s identity and allowed callback URLs.	Prefer this when your client supports published client metadata. In Claude, this is Use Claude’s published identity.
DCR	The client registers through the advertised registration endpoint and receives its own OAuth client credentials.	Use this when your client supports automatic registration. In Claude, this is Register automatically.

Both methods still require user authorization. Your client handles identification or registration during setup; you do not need to create an OAuth client manually for either method.

For client implementers

Use MCP OAuth discovery to obtain the authorization server’s endpoints and supported capabilities. The public discovery documents are:

The authorization server advertises client_id_metadata_document_supported: true for CIMD and a registration_endpoint for DCR. The CIMD profile supports public clients using token endpoint authentication method none. DCR supports public clients using none and confidential clients using client_secret_basic. Here, none means the OAuth client has no client secret; user authorization is still required. Use PKCE S256 and preserve the exact server URL as the OAuth resource value. Request offline_access if your client needs refresh tokens. See the MCP authorization specification for discovery and registration behavior.

Available tools

Tool	Purpose
`get_corpus_overview`	Describe the collection’s available material and coverage.
`retrieve_evidence_context`	Retrieve evidence and answer instructions for a question.

The overview helps with coverage questions. You can call retrieval directly for an evidence question. A successful empty retrieval confirms connectivity but does not establish support for an answer. Preserve the returned evidence instructions, source attributions and citation URLs when passing context to a model. See the citation guide for claim attribution, link expiry and missing-link behavior.

If something goes wrong

Problem	What to check
The client cannot connect	It supports Streamable HTTP and uses the exact server URL above.
OAuth setup fails	The client supports discovery, PKCE S256, and CIMD or DCR. Check its visible error and registration settings.
Sign-in or consent is denied	Your AI Baseline API key grants access to this collection.
Tools are missing after authorization	The connection is enabled and the client has refreshed its tool list.

Get Started

Connect

Guides

API Reference

Select a knowledge base

Connection settings

Add and authorize the connection

Client registration

For client implementers

Available tools

If something goes wrong

​Select a knowledge base

​Connection settings

​Add and authorize the connection

​Client registration

​For client implementers

​Available tools

​If something goes wrong

Select a knowledge base

Connection settings

Add and authorize the connection

Client registration

For client implementers

Available tools

If something goes wrong