Expandi ProductAI agent connection
Designed how a salesperson hands an AI agent the keys to their Expandi account, safely.
Expandi opened its product to Claude and ChatGPT over MCP. The tools existed; what didn't was a way for a non-technical salesperson to connect one, and to understand what they were agreeing to before an agent could message their leads.
- AI clients at launch: Claude and ChatGPT
- 2
- permissions, listed in plain language
- 6
- prompt to copy to set it up
- 1
- The first draft of setup was written for developers: a server URL, a token, a config file. Our users are SDRs.
- Consent was the bigger risk. Any app can call itself anything, so a fake "Claude" could ask for the same access as the real one.
- Sales teams already using an AI assistant every day, connecting it without IT.
- Developers building custom agents were the smaller group. They got documentation, not the main flow.
- One prompt to copy instead of configuration: the assistant walks the user through its own setup.
- Trust is earned per app. Known clients get a verified badge; unknown ones lose the logo, get an Unverified tag and a warning, and still can't hide what they'll be allowed to do.
- A setup page per tool, with numbered steps and one prompt to copy.
- A consent screen listing six permissions in plain language, and where to disconnect.
- Anything an agent writes lands as a draft. A person sends it.
- Shipped July 2026. A salesperson connects an assistant alone, in about two minutes, and around 600 accounts connected one in the first month.
- No message reaches a lead without a human pressing send.
(Keep browsing)
