Platform guide

🔌 Integrations & secrets

6 min read · Updated

Integrations (Settings → Integrations) connect the tools your business runs on - mail, calendars, project trackers, analytics, payments, CRMs and more. Each connection supports multiple accounts, a primary, renames, and scope presets: connect read-only where read-only is enough. Per-agent access lives one click deeper: Manage on a connected account can pin an agent to a narrower level or block it from the app entirely - enforced on every call, no restart needed.

Every account row shows its health (expired credentials get a Reconnect), and “last used” jumps to that integration’s runs - so “what has it been doing with my calendar?” always has a receipt.

Custom integrations (Settings → Custom integrations) hold credentials for anything we don’t natively support - grouped fields, write-only values, masked display. Agents can use them; nobody can read them back out.

Recommended is the first row of the catalog: products we vouch for, with dedicated agent tools. SiteGPT (AI support chatbot: sign in and approve), Revid (short-form video: paste your API key), PostSyncer (social scheduling: paste your API token), and Bazzly (Reddit marketing: paste your API key) are live today. Connect from the card and the tools reach your agents within a couple of minutes. Agents can never buy anything in these products. Revid renders and Bazzly reply generation spend the credits already in your Revid or Bazzly account (agents estimate Revid costs up front and flag you when a balance runs low); buying more is always yours to do, in the vendor’s own dashboard.

Add any MCP server

Any MCP server connects from the same catalog: Integrations → Custom → Add MCP server (also reachable from Settings → Custom integrations). Paste either the server's https URL or the JSON config from its docs. The exact "name": { ... } snippet vendors publish works as-is, trailing comma and all, for hosted servers and command-based ones alike.

If the config carries a key or token, it lands in the credential vault like every custom credential: write-only, masked, usable by agents but never readable back. Your agents get every tool the server exposes, named server_tool, within a minute or two of connecting.

Pasting the config to an agent in chat works too: the agent sets the server up and reports back which tools arrived. Prefer the dashboard when the key doesn't exist yet, so fresh secrets stay out of chat history.

Two guardrails to know: a server can't take the name of a built-in integration or an existing custom credential (rename it if the connect says so), and two different endpoints never silently merge into one entry.

⚠️ Heads up: Scope discipline beats trust: give the narrowest preset that does the job, and expand only when a real task needs it. You can change scopes any time - agents adapt on their next run.