Some work starts as a question and turns into several next steps. A Canopy bot keeps that conversation together. It can gather context, start a coding thread and bring a decision back to you.
Give each bot a clear responsibility
Create a bot for a recurring responsibility, such as project follow-up or your morning brief. Choose its instructions, model and computer. Each bot has one ongoing conversation, so earlier requests and later updates share a place.
A useful instruction names the sources, the decision and the point where you want to hear back. For example, ask for unresolved checkout issues that need your review. Avoid an open-ended instruction to handle everything across every account.
Bring connected tools into the conversation
Grant a bot selected connected accounts from Settings → Connectors. You can keep work and personal accounts separate, including multiple accounts for one service. The bot receives tools from its granted accounts, rather than every account in your installation.
A project bot might read Linear issues and GitHub activity before summarizing a blocker. Account access and action approval are separate choices. Permission to use an account does not automatically approve every write the external service offers.
Turn a conversation into coding work
When an issue needs code, the bot can start a Thread in a chosen project workspace. It sends the task to an available coding agent, follows completion and can send further instructions. Coding happens on the selected worker.
You can open the Thread to inspect its history and changes. Approval requests from Threads started by the bot also appear in its conversation. Choose standing approval rules for familiar actions while keeping protected actions under your control.
Set it up for the way you work
Start with the desktop app or your own Canopy server. Connect Claude Code, Codex or Grok on the assigned worker, then create a bot. Add one connected account, try a specific request, then introduce a routine once its scope is clear.
Bots use official harnesses on workers and do not directly edit project files or run local shell commands. Their assigned worker must be available. Conversation history is retained, but model context remains finite; treat important ongoing instructions as explicit bot settings.