How to Connect Tools to an AI Agent with Arcade

November 21, 2026

Arcade is infrastructure for giving AI agents real tool access. Instead of you storing OAuth tokens for Gmail, Slack or HubSpot and pasting them into a prompt layer, Arcade holds the connection per end user and the agent asks for permission at the moment it needs it.

If you are building anything that takes an action on behalf of a person, this is the layer between your agent loop and the outside world.

What problem does Arcade actually solve?

Three, stacked. Tool calling: it provides pre-built tool definitions with typed parameters so the model does not invent an argument shape. Authentication: it stores per-user OAuth and refreshes it. Authorisation: it shows the user a consent screen before a sensitive call runs.

The last one matters most in production. A read-only calendar tool can run silently. Sending an email cannot.

How does auth delegation work for end users?

Each user authenticates once against the provider through Arcade’s flow. From then on, a tool call includes the user’s identity and Arcade injects the matching credentials. Your backend never sees the token.

That design means a support agent in a shared Slack workspace cannot read another customer’s mailbox through the same tool — the credentials are keyed to the person, not the app.

Practical consequence: your agent’s prompts must carry a stable user ID. If you key sessions by a random chat ID, auth lookups fail and you get empty results that look like the tool is broken.

What does a first working setup look like?

  1. Pick one tool with low blast radius: list calendar events, read a Slack channel, fetch a Jira issue.
  2. Define or import the tool and confirm the parameter schema matches the provider’s API.
  3. Run the auth flow with your own account first and watch the consent screen.
  4. Wire the tool into your agent loop and log every call, including the returned payload size.
  5. Add a destructive tool (send, create, delete) only after the read path is stable.
  6. Add human-in-the-loop confirmation on anything that writes.

That sequence takes an afternoon for a read tool and a day or two for a writing tool with consent.

What are the sharp edges?

The consent UI interrupts long conversations. If your agent asks for permission mid-answer, some products handle it poorly — you need a queued “waiting for approval” state or the user sees a hung response.

Rate limits belong to the underlying provider. Arcade can pass a rate-limit error through; your agent needs to recognise it and back off rather than retry three times into a 429.

Payload size. Gmail search returns message bodies by default more often than you expect; trim fields before they hit the context window.

Who is this for?

Product teams building an agent feature where end users see it. Multi-tenant SaaS where per-user permission is a legal requirement. Developers who have already built a raw tool-calling loop and are tired of maintaining OAuth refresh code.

It is not for a personal automation script where you are the only user — a hardcoded token in an environment variable is simpler. It is also not a workflow builder: if you want a visual canvas with branching steps, that is a different category of product.

Quick Summary

  • Arcade provides tool definitions, per-user OAuth and consent prompts for AI agents.
  • Credentials stay with Arcade; your backend never stores provider tokens.
  • Sessions must be keyed to a stable user ID or auth lookups silently fail.
  • Start with read-only tools, add write tools plus confirmation after.
  • Trim provider payloads before they hit the model’s context window.
Tags: