Nango has written the clearest explanation of MCP elicitation I have found. In MCP elicitation: how agents ask users for input mid-tool-call, Emmanuel Oyibo lays out the mechanism by which a server pauses inside a tool call and asks the user for what it is missing, instead of demanding every detail up front or letting the model guess. The division of labor is one sentence: “The server controls what to ask and when. The client controls how the question appears.” The spec history is precise. Form mode arrived in 2025-06-18, URL mode in 2025-11-25, and “the 2026-07-28 specification supports both modes,” the same revision that deprecated sampling. The rule that matters most is the prohibition: “Servers must not use form mode to collect secrets or credentials used to access services or authorize transactions: passwords, API keys, access tokens, and payment credentials.” Those go through URL mode, which is where OAuth lives.
The one statistic in the post is borrowed rather than claimed. “Anthropic reported that Claude Code users approved roughly 93% of permission prompts, and paid less attention as the number of prompts grew,” which is the case for asking rarely: “Reserve confirmation for actions whose consequences justify interrupting the user.” The vendor figures are the ones on the Nango pitch, 1,000-plus APIs and 7,000-plus tools, and they are worth flagging because the catalog record still says 400 APIs and 600-plus integrations. One of those numbers is stale, and it is probably ours. What survives independent of the marketing is the SDK survey, which documents where elicitation is supported as of September 2026: form and URL modes in the TypeScript and Python SDKs, form only in the Vercel AI SDK, and nothing yet in the OpenAI Agents SDK or the Anthropic MCP connector.
The catalog maps the post’s advice onto Nango’s own surface directly. The Nango provider page lists 6 API pages. URL-mode elicitation, where a server hands the user an OAuth flow rather than asking for a token, is the Nango Connections API, which creates and manages those authorized connections. The tool call that resumes once the credential exists runs through the Nango Proxy API or the Nango Actions API. The MCP server dimension is lit on the Agent Readiness score, so the catalog agrees Nango ships the thing the post is describing. The agentic access profile maps 21 operations, 14 of them acting.
The Kin Score is 44.5, developing band. Discoverability carries it at 66.7, developer ergonomics at 52.4, contract quality at 51.0. Contract governance is 0.0, the only zero facet among today’s three providers. The Agent Readiness score is 28.4, agent-aware, and the unlit dimensions are the ones the post is about. Consent identity is unlit. Delegated identity is unlit. Dynamic client registration is unlit. Elicitation is the protocol’s answer to consent mid-action, and Nango has explained it better than the specification does, but the catalog cannot find a description of how a user consents to an agent acting through Nango’s own API. The post tells servers to ask the user at the right moment. The record does not yet say how Nango’s would.