Field noteNov 16, 20266 min read1 source

MCP: AI/ML API Documentation

In one minute

AI/ML API’s MCP guide explains how to connect its model catalog and inference services to Claude Desktop, Claude in a browser, Cursor, Claude Code, and other compatible clients. OAuth lets builders authenticate in a browser rather than paste an API key, then use AI/ML API tools inside their existing client; calls remain billable to their AI/ML API account.

What happened

AI/ML API’s documentation describes a production remote MCP server at https://mcp.aimlapi.com/mcp. Its intended clients include Claude Desktop, Claude on the web, Cursor, Claude Code, and other applications supporting remote MCP over Streamable HTTP with OAuth. The guide says that a client can connect using OAuth 2.1 with PKCE and dynamic client registration. It also supports a separate API-key route using an Authorization: Bearer header; the page’s setup instructions focus on browser-based OAuth.

In the OAuth flow, the client contacts the server, discovers the authorization server, registers itself, and opens a browser. The user signs in to AI/ML API and approves access; the client receives and stores a token, then refreshes it. The guide says users do not handle an API key in this flow. It cautions that the computer’s clock must be correct because OAuth tokens are time-sensitive.

The page provides client-specific setup instructions. In Claude Desktop, users add a custom connector and then click Connect to complete browser authentication. Claude on the web uses a similar connector flow, although the company notes that custom-connector availability depends on the Claude plan; on Team and Enterprise plans, an organization owner handles setup. Cursor users add an mcpServers entry containing the server URL to a global or project-level mcp.json file, then enable and log in to the server in settings. The guide explicitly says to omit a headers block to trigger OAuth. In Claude Code, the documented command adds the server over HTTP; by default it is available only in the current project, while --scope user makes it available everywhere. The user then authenticates from an interactive session.

Once connected, AI/ML API says users can search and compare models, run LLM chat and embeddings, generate images, video, and audio, manage long-running generation jobs, and check account balance and usage. The examples include comparing models by price and context window, asking for a cheaper model for a recent task, and submitting a generation request. The page describes AI/ML API as an aggregator of more than 1,000 models. Every call is billed to the user’s AI/ML API account, on the same basis as its REST API; the guide directs users to top up their account before billable inference.

The guide also covers disconnecting and troubleshooting. Removing a client connector may not revoke its OAuth token on AI/ML API’s side; a fully clean reinstall requires revoking the app under account connections before reconnecting. It gives distinct removal steps for the supported clients, and notes that in Cursor the configuration must be deleted from the JSON file rather than only hidden in the interface.

What the archive adds

We have not covered this before; it opens a new thread.

Trend signal

This is a concrete example of an agent protocol serving as an integration layer between a client and an external tool provider. MCP lets supported clients reach model discovery, inference, and account tools through one server endpoint, rather than requiring a separate bespoke connector for each client. The example supports New Runtime’s agent protocols become the interoperability layer pattern, while leaving the quality and portability of the integration to be tested in practice.

Why now

The guide’s relevance comes from its support for several established AI clients, including coding environments and Claude clients, through a shared remote MCP endpoint. That gives teams already using those clients a documented way to try an aggregator from inside their working interface. The page says it was updated recently, but does not explain what prompted the update or establish a wider market timing story.

Why it matters

For teams that compare providers or use different models for different tasks, placing catalog search and inference inside a familiar client can reduce the friction of trying alternatives. The same connection also exposes account balance and usage, so an operator can ask cost-oriented questions without leaving the workflow. That is useful only if teams keep the billing boundary clear: the client is an interface to AI/ML API, not a substitute for its account, budget, or usage controls.

OAuth avoids placing a long-lived API key in a client configuration, but it does not remove the need to manage authorization. Teams should understand which account is being authorized, where the client stores tokens, who can use a user-scoped connection, and how to revoke access. The documentation’s distinction between adding a connector and actually authenticating it is especially practical: a configured server that has not completed sign-in will not work. For a team evaluating MCP integrations, this guide is a useful setup reference and a reminder to test permissions, billing, and removal procedures before broader adoption.

Limits and uncertainty

The guide describes how the connection is intended to work; it does not provide independent measurements of latency, reliability, model-selection quality, or performance across the listed clients. It also does not establish whether every model or modality in the catalog is available through every tool at all times. Usage is billed to the AI/ML API account, but the page does not give a single price for inference because costs depend on the chosen model and operation. OAuth reduces the need to paste a key into client configuration, but the documentation alone does not demonstrate token-storage security or the organization-level controls available for every client. Builders should verify current client-plan requirements, account permissions, and billing before relying on the integration in production.

Skill to build now

Try the integration in a non-production client or project. Add only the MCP URL, complete browser authentication, and first ask the client to check the account balance. Then test one low-cost model-discovery request and confirm where usage appears in the AI/ML API account. Record the exact revoke and removal steps for your client before teammates connect their own accounts; for Cursor, include deleting the JSON entry, and for a clean OAuth reset, revoke the authorized app as well.

Reader takeaway

AI/ML API’s MCP server brings model discovery, inference, jobs, and account checks into compatible AI clients, with OAuth as an alternative to pasting an API key. The integration is immediately testable, but it does not change who bills the requests: usage remains tied to the AI/ML API account. Treat authentication, scope, revocation, and cost checks as part of setup—not cleanup.

Evidence / sources

  1. 1docs.aimlapi.com/quickstart/mcp
All Field Notes

Next reads

Read next

Open the graph
  1. 01topicClaudeExplore the claude topic hub.
  2. 02topicMcpExplore the mcp topic hub.
  3. 03topicClaude CodeExplore the claude-code topic hub.
  4. 04archiveField NotesOpen the latest editorial analysis.
  5. 05sourcesSourcesSee the original sources this archive cites.
Who read this page?Machine requests, hidden until opened

Loading the privacy-safe route aggregate...

Open the JSON contract