One ask on the left. On the right, connections then the editor then the dashboard, cutting between each other as two systems get credentials and thirteen automations come up: order-to-cash out to Workday, finance state back into HubSpot.
Settings
Tell Claude
You can comment on one or more steps across multiple automations. When done, return to your Claude Code session and run /apply-comments to have Claude read all of your comments.
The apps and accounts you have connected.
https://app.apiant.ai/oauth/callbackWhat is captured. The connections screen is reproduced from the live App Connections page at app.apiant.ai: the card grid, the Connect App dialog and its type-ahead over the catalogue, the NEEDS YOUR KEYS badge that marks a bring-your-own-keys app, the client_id and client_secret fields that open inline in the same dialog, the exact redirect URI the vendor has to register, and the Continue to handoff. Authorization happens on the vendor's own site, and APIANT never draws a consent screen, so neither does this replay. The editor and dashboard chrome come from the same app: palette, pill nodes, nested true/false lanes, parallel branches, the loop rail, the step settings panel with its Tell Claude box, and the dashboard's row layout, usage card and sort bar. Every number in the meters is counted off what is actually on screen.
What is illustrative. The whole HubSpot / Workday suite is a worked example. It is not a customer integration, it was never built, and it was never run. The Workday tenant name and client id are stand-ins, no credential was exchanged, and the transcript is a reconstruction using APIANT's real MCP tool names. Field names follow each vendor's public API naming. Because no run happened, no task counts or timings are claimed; the dashboard shows them unset, which is also how the live one reads before a first run.