Documentation by ClonePartner

Start typing to search.

Using Envoy

AI assistant

View Markdown

Open the AI Assistant sidebar, attach context and files, run slash commands, and approve or reject destructive actions.

AI assistant

The AI Assistant is a global sidebar panel available on every authenticated screen — it is not a separate page. Operators use it to ask questions about Tasks, Runs, and Connectors, search the app, and perform API actions with inline approval for anything destructive.

Open the assistant

  • Select AI Assistant at the bottom of the left navigation.
  • Press ⌘/Ctrl-Shift-A to toggle the panel from anywhere.
  • On phones and tablets, tap the brain icon in the top bar (Open AI assistant).
  • Append ?ai=<conversation_id> to any authenticated URL to deep-link a specific conversation.

⌘/Ctrl-K opens the command palette (app search), which is a separate surface from the assistant.

Screen overview

The panel has three areas:

  • Header — the current conversation title (or New chat) acts as a switcher trigger. Next to it sit icon buttons: Rename, New chat, Delete conversation (or Clear for an unsaved chat), and Close.
  • Message list — your messages, the assistant's replies, collapsible "Thought for …" reasoning rows, and collapsible tool-activity rows (for example Read task, Searched app, API call). Expanding a tool row shows its Request and Response. Records the assistant mentions appear as chips that link to the matching page.
  • Composer — a single rounded input at the bottom.

The composer

Type a message and press Enter to send (Shift+Enter inserts a newline). The placeholder reads Ask Envoy AI… · / for commands.

  • Attach files — the paperclip button, pasting an image, or dragging files onto the composer uploads attachments. Images (PNG, JPEG, GIF, WebP) and text files such as .csv, .json, .yaml, .log, and .md are accepted, up to 5 attachments per message.
  • Mention records — type @ to search Tasks, Jobs, Connectors, and other records and attach them as context chips.
  • Model picker — when more than one model is available, a model button next to the paperclip lets you switch; the default is marked · default.
  • Send / Stop — the filled send button becomes a solid square Stop button while the assistant is responding.
  • Context meter — a circular gauge shows how much of the model's context window the conversation uses.

If the panel shows AI is not configured., the server has no AI provider key — see the configuration note below.

Slash commands

Type / to open the command menu:

Command What it does
/summarize Recap this conversation — findings, decisions, next steps
/explain Explain what I'm currently looking at
/diagnose Investigate a failure and suggest a fix
/task Scaffold a new task from a description
/find Search tasks, jobs, connectors, runs, triggers
/docs Explain a step, connector, or concept from the docs
/review Review the current task YAML for issues
/new Start a fresh conversation
/rename Rename this conversation

What the assistant can and cannot do

The assistant reads freely but pauses for approval before it changes anything sensitive:

  • Read — it lists and reads Tasks and reference docs, searches the app, samples Connector data, describes API operations, and navigates you to pages. Connector configurations are redacted before the assistant sees them.
  • Change — it calls the REST API as you, with your role's permissions. Destructive API calls (including user invites) pause with an approval card.
  • Never — authentication, API-key management, executor administration, and nested chat endpoints are blocked entirely.

Approve or reject an action

When a destructive call is proposed, a card labeled Destructive / Approval required shows a summary plus the exact method and path. Choose:

  1. Approve — runs this one call.
  2. Approve always (when offered) — runs the call and skips future prompts for the same action in this chat.
  3. Reject — tells the assistant no; the conversation continues.

A pending approval is stored server-side and expires after 15 minutes. Sensitive action summaries are audited with secret fields redacted.

Manage conversations

Conversations are saved automatically. Select the title in the header to open the switcher: a Search chat history… field on top, a New chat entry, and your recent conversations beneath it. Rename and Delete conversation live in the header; deleting permanently removes the conversation and all its messages.

Long conversations are summarized automatically — an expandable Earlier conversation summarized divider marks the compacted history, and the full transcript stays visible. If a response ends abnormally (provider error, stopped stream, or a limit reached), the chat shows the reason with a Continue or Retry button that resumes without repeating completed work.

Configuration and knowledge

Enabling the assistant and choosing models is a server-side ai configuration concern — see Server configuration. An admin can also store the provider API key from the Settings screen — see Users, roles, and settings.

The assistant's knowledge base is Envoy's built-in reference corpus. Your deployment can override or extend it with AI references; enabled references shadow the built-in document with the same name.