The assistant

The assistant is a chat that knows about your workspace. It can look things up across your integrations, explain what it found, and, under the configured approval policy, make changes.

Where to use it#

• Dashboard: open Assistant.
• Phone app: the Chat tab, with a list of your chats in the drawer.
• Desktop app: the Escanor button on every screen (or Alt+J), and the AI section.
• By voice: see Voice.

Conversations are saved until you delete them. You can rename, pin, archive and delete them. Task runs are kept for one year.

What it can do#

• Answer questions using live data from your connected tools.
• Summarise deployments, incidents and activity.
• Prepare changes (open a pull request, redeploy, restart) and ask you to approve them.
• Read files, images and PDFs you attach.

It reports each step while it works. Its answers are labelled as AI-generated: they can be wrong or out of date, so check them before you rely on them or let them change a production system.

Approvals#

Actions that require approval under the workspace or execution policy pause for review. In the apps, approve or deny from the chat or Approvals list. Inspect the proposed operation, provider and arguments before approving. Read-only lookups generally proceed without a prompt. Provider confirmation checks, assistant approvals and machine execution policies are separate controls; do not assume every possible write always produces the same dialog. See Permissions and approvals.

Attachments#

Use the + button in the message box to attach photos, PDFs and text or code files (up to four files, with size limits shown if you go over). Photos are shrunk before sending. Word and Excel files are not supported yet: save them as PDF first.

Which model answers#

The assistant has an Auto mode. It looks at each message (small talk, a quick lookup, code, a multi-step task, hard reasoning) and picks a model suited to that work, then falls back to another if one is slow or unavailable. You can also choose a specific model; availability and fallback behavior depend on the selected routing mode. Check the model shown with the response and any error before relying on a specific model.

The workspace assistant uses cloud models. The desktop app also has a separate local model playground backed by Ollama; loading a local model there does not replace cloud routing throughout the assistant. When a model struggles, Escanor skips it for a few minutes instead of retrying it on every message.

Speed#

Short, conversational messages are answered by a fast model straight away. Messages that need tools go to a more capable model, and the answer arrives when the work is done. If a model takes too long, the next one is tried in parallel rather than waiting for the first to fail.

Limits#

The number of assistant runs per month depends on your plan: 50 on Free, 1,000 on Pro, 10,000 on Team. Settings > Usage shows how many you have used. See Plans and billing.

When it cannot help#

• "Not connected": connect the integration it needs (see Integrations).
• "Needs reconnect": reconnect it.
• A model error: try again, or pick another model from the list shown.

A practical first conversation#

Start with a connected provider you can safely read: “List my recent deployments and explain which failed. Do not make changes.” Verify that the response names the provider and concrete resources you recognize. If it asks to connect a provider, complete that step in Integrations and retry the read.

For a change, specify the project, environment and intended outcome. Ask for a plan before execution: “Explain how you would recover the failed staging deployment; wait for me before rolling back.” Review approval cards; the assistant's prose is not evidence that an operation completed. Inspect the provider's resulting state after an approved action.

Recovery and privacy#

If a response stalls, check the activity/run state before sending the same write again. An interrupted response does not prove the provider operation was cancelled. Deny an unexpected pending request and use the run's stop control where available. For an expired approval, create a new request after checking current state.

Attachments are sent to the AI model that answers. Do not attach credentials or private material you do not want sent to it. Escanor does not train AI models on your messages, attachments or workspace content. Cloud connectivity, provider permissions, plan quotas and model availability affect execution. See Troubleshooting and Account and sessions.