# The APIANT assistant

> The APIANT assistant builds, edits, tests and diagnoses automations from a conversation in the app, costs 20 credits per message, and hands connector work to Claude Code.

The APIANT assistant is a conversation in the app, under **Assistant**, that builds and changes automations on your account by calling a subset of the APIANT tools Claude uses over MCP. You describe what you want; it looks up apps in the catalog, asks you to connect them, builds, tests and reports what happened. Each message you send costs 20 [assistant credits](/docs/assistant/credits).

The Assistant page lists your conversations, shows your account usage, and links to Connect to Claude Code.

## Where the assistant appears

| Entry point | What it opens |
|---|---|
| **Assistant** in the top navigation | Your conversation list and a new conversation |
| **Ask about this** in an automation's **More actions** menu on the dashboard | The conversation about that automation, created if it does not exist |
| **Ask about this** on a failed run in the automation's execution history | The same conversation, opened with a question about that run |

These entry points show only when the assistant is available on your account and you have credits left for the month.

## What it can do

- **Build an automation** from a description, including forms, chat widgets, web services, fan-out and fan-in, delays, approvals, two-way sync, and automations exposed as MCP tools.
- **Connect apps** by posting a **Connect** button when a build needs an app you have not connected. OAuth sign-in happens on a separate page and returns you to the conversation.
- **Edit an automation**: change a mapping, add or remove a step, rename it, move it to a folder.
- **Apply step comments** you left on an automation's diagram.
- **Test** by running the automation, with sample trigger data when needed, and report each step's result.
- **Diagnose and retry** failed runs.
- **Turn automations on and off.**

## What it hands to Claude Code

The assistant works only through its tools, with no terminal or browser control. It does not:

- Build or change connectors, triggers, actions or assemblies, or integrate an app the catalog does not have.
- Publish or deploy automations to other accounts.
- Tune how error alerts are worded, suppressed or routed.
- Replay a recorded webhook, restart a run from a step, or edit a run's data to force a branch.
- Register an OAuth app on a vendor's developer portal for you.

When a request needs connector work or a missing app, the assistant says so and posts a **Connect to Claude Code** card. Its button opens the same install dialog as the account menu. Claude Code uses the same account and no credits. See [Connect Claude Code](/docs/mcp/claude-code).

## How a conversation works

- **Progress.** While a run is in progress, a line next to the spinner says what the assistant is doing. **Stop** ends the run.
- **Queued messages.** A message you send during a run is marked queued and picked up at the next step.
- **Leaving the page.** **Notify me when done**, next to **Stop**, asks your browser to notify you when the run finishes or stops early. The browser notification needs the tab left open, so tick **Email me too** if you might close it. The notice is sent once, for that run.
- **One automation per conversation.** A conversation is bound to the first automation saved in it, or to the automation you opened it from with **Ask about this**. When another save happens in a bound conversation, the assistant posts "I've kept this conversation on the first automation" and suggests a new conversation. The save itself still happened.
- **Failure notices.** When an automation bound to a conversation fails, the conversation gets a message about the failure. If you were not on the page to read it, you get an email with the same words. The same failure repeating is not announced on every run.
- **Managing conversations.** Each conversation's menu has **Rename**, **Archive** and **Delete**. Deleting a conversation does not delete the automation it built.

## Credits

Each message you send costs 20 credits when the assistant picks it up, however many steps the assistant takes to answer it. A message refused for lack of credits is not charged. Reading a conversation is free. Your plan's monthly allowance and any credit packs are shown in the **Account usage** panel. See [Assistant credits](/docs/assistant/credits).

## Next steps

- [Build and edit with the assistant](/docs/assistant/build-and-edit)
- [Ask about an automation or a failed run](/docs/assistant/ask-about-a-run)
- [Assistant credits](/docs/assistant/credits)
- [APIANT over MCP](/docs/mcp)