Blog / Onboard agents through MCP

Agent onboarding for SaaS: OAuth gets the agent in, it doesn't tell it what to do

Most “agent onboarding” advice stops at authentication. That part is now standard. The harder part is the first minute after: does the agent know what your product is for, where the user left off, and what to call first?

Leonid Bugaev · Published 21 Oct 2026

Getting in: standard MCP OAuth

The MCP spec covers this: MCP authorization. A compliance SaaS MCP server we studied does it by the book:

  1. An unauthenticated request gets 401 with a WWW-Authenticate header pointing to the protected-resource metadata.
  2. That names an authorization server with dynamic client registration, PKCE (S256) and refresh tokens.
  3. The client registers itself, a browser opens, the user signs in, and the connection takes the user's role.

One detail worth copying: its 401 text is written for the agent, telling it to clear its tokens and reconnect, instead of a bare “unauthorized”.

After the login

The same server then sent 115,077 characters of instructions (26,633 tokens). Its most important rule, “call the status tool first on every connection”, sat at character 28,484.

Instructions sent
115,077 chars
Where the first-call rule was
char 28,484
What one app showed the model
first 2,067 chars

To its credit, it repeats the rule in tool descriptions, which is how agents still found it. Lengths and cut-offs per app: server instructions length.

Want to see your onboarding through an agent's eyes?

We connect the agent apps your users run to your MCP server as a new user would, and show what the agent sees and does in its first calls.

Test your onboarding

The first-call recipe

  1. Instructions start with the rule: what the product is for, then “call get_instructions (or your status tool) first”. Under ~2,000 characters.
  2. A start-here tool with a tiny schema that returns the user's state and one next step.
  3. One line in other tools' descriptions: “Before your first call, call get_instructions once.”
Claude Code, tool only
1/24 (1–20%)
Claude Code, full recipe
24/24 (86–100%)
Codex, tool only
9/24 (21–57%)
Codex, full recipe
24/24 (86–100%)

Runs that called the start-here tool first, picked the right job and followed step 1. Details: make agents read your instructions first.

What stays on the web

Go deeper: When the agent doesn't recognise your server, Agent skills vs MCP server, Next-step hints.

Back to onboarding agents through MCP.

How we measured

FAQ

How do I onboard AI agents to a SaaS product via MCP?

Use standard MCP OAuth so clients can register and users sign in, then make the first minute work: short server instructions that start with what to call first, a start-here tool that returns the user's state and the next step, and clear marking of tools that spend or contact people.

Is OAuth enough for agent onboarding?

It gets the agent connected with the user's permissions. It doesn't tell the agent what your product does or where to start. In our tests that's where agents went wrong.

Where should the first-call rule go?

At the very start of the server instructions, in the start-here tool's description, and as one line in other tools' descriptions, because apps cut instructions (Claude Code at 4,096 characters, Grok Build around 2,048).

Want to see your onboarding through an agent's eyes?

We connect the agent apps your users run to your MCP server as a new user would, and show what the agent sees and does in its first calls.

Test your onboarding