Assistant

Ask about your apps in plain words: why a deploy failed, why an app is slow, what changed, what to do next.

The assistant is in beta. It only works once the platform has been set up for it; until then it answers "not set up yet" (see assistant_unavailable).

Open Assistant in the dashboard, or run deplane ask "QUESTION". On a failed build, Ask the assistant why opens it with the question filled in.

What it can see

The assistant reads what the dashboard shows, in your team only: your apps and their status, recent app and build logs, releases, metrics, alerts, custom domains, workers and cron jobs, databases, Redis, MySQL and ClickHouse, your variables' names, and the GitHub connection.

It never sees:

• variable values. Before anything reaches the model, every value of your team's variables and the passwords of its databases are replaced with [REDACTED], wherever they appear, logs included. Values shorter than 8 characters are replaced only where they follow their name (DEBUG=1).
• commands and their output. Release commands, workers' and cron jobs' commands, deplane run commands, and the output of release commands and runs can hold secrets. It sees whether they succeeded, their exit codes and how long they ran.

What it can do

The assistant acts only on the app the question is about: asked from that app's page in the dashboard, or with deplane ask --app NAME (or in the app's folder). That includes the app's own databases, Redis, MySQL and ClickHouse. A question about no app gets answers, not changes. It never creates an app.

It makes ordinary changes itself, as you, with your team's limits: redeploying, deploying the connected GitHub branch, raising instances or size, adding a custom domain, a disk, or a data store for the app, and backups. Each is in the team's activity log, marked as done by the assistant for you.

It asks first, with a card you confirm, before anything that runs new code, takes something away or costs money:

• anything that runs new code with the app's variables: deploying an image, adding a worker or cron job, changing a command;
• lowering instances, size or the autoscaling ceiling, turning sleep when idle on, stopping or shrinking a worker, rolling back;
• removing variables, and detaching a database, Redis, MySQL or ClickHouse (the data store is kept);
• setting variables: it names them, and you type the values into the card, so it never sees them;
• removing a domain, the private network, deploying the latest commit;
• restoring a database into a new one and switching the app to it (the original is not changed);
• deleting the app, or one of its data stores. These cannot be undone: the card works only once you type the name.

A card runs the change as you, with the same passkey check as anywhere else. Anything that costs extra will ask first too; nothing does today. On Free it knows your plan's rules: sleep after 5 or 15 minutes only, cron jobs but no workers, and no private network (limit_sleep).

It cannot start deplane run, since a run has the app's secrets and does whatever its command says: it tells you the command to type.

In the CLI

deplane ask behaves the same. For a change that needs your confirmation it asks Run this? [y/N], or for a delete asks you to type the name, then runs the deplane command; --yes never skips these. Without a terminal (CI, a script, an AI agent), nobody can confirm, so it changes nothing at all: if the answer needs any change, it exits 1 (assistant_needs_confirmation) and lists the commands to run.

Where your data goes

Answers come from Claude, made by Anthropic. To answer, Deplane sends Claude your question and the data it looks up for it (for example log lines and metrics). Anthropic does not use API data to train its models. Text inside your apps' logs is treated as data, never as instructions to the assistant.

Answers can be wrong: check them before you act. You can ask 40 questions an hour.

This page is also available as Markdown. AI agents can start at /llms.txt.