PerchAI
Using Perch

Ask, Plan, and Agents

The three chat modes decide whether a turn only answers, drafts a plan for your approval, or does the work. They are separate from the permission mode, which decides how much runs without asking.

Every turn in Perch Terminal and the CLI runs in one of three chat modes. The mode is the first thing to reach for when Perch is doing more, or less, than you want.

ModeWhat a turn does
AskAnswers from the chat alone. No tools, and nothing changes.
PlanLooks around read-only, drafts a plan, and waits for your approval before changing anything.
AgentsDoes the work: reads and edits files, runs commands, and uses connectors, as your permission mode allows.

Switch in the composer's mode menu, or in the CLI with /mode ask, /mode plan or /mode agents. Bare /mode shows the current one.

Ask

Ask is for talking about the work rather than doing it. Perch answers from the conversation, with no tool calls at all, so nothing is read, written, or run. Use it to think out loud, to ask what an approach would involve, or when you want a quick answer and not a flurry of file reads.

Plan

Plan is the mode for work you want to look at before it happens. Perch inspects the workspace, reads files, searches, and then writes a plan. Nothing that changes anything runs while it plans: every mutating tool stays blocked, so you can read the plan knowing your files are untouched.

What you get is a plan card with a title, the objective, and the steps. On desktop the card sits in the transcript with approve and reject under it. In the CLI it prints and asks:

Plan ready for approval: Fix the failing auth tests
Approve and execute this plan? [y]es / [n]o:

The plan is also written to your folder, as markdown under .perch/plans/, so you can read it in your editor, keep it, or hand it to someone else.

  • Approve and Perch starts executing. The chat mode goes back to Agents for the work.
  • Reject and nothing runs. Say what to change and Perch revises the plan.

Plan mode is the right default for a large refactor, anything touching files you care about, or the first task in an unfamiliar repository.

Plan is also a permission mode

Confusingly, Plan appears twice: as a chat mode (/mode plan) and as a permission mode (/permission plan). They do the same thing from different directions, and setting either one keeps mutating tools blocked until a plan is approved. The practical difference is that the permission mode sticks until you change it.

Agents

Agents is the working mode and the one most turns run in. Perch reads, edits, runs commands, calls connectors, spawns helpers, and drives the browser. How much of that happens without asking you is not the chat mode's business: that is the permission mode.

Chat mode and permission mode are different settings

This is the distinction worth internalizing:

  • Chat mode decides what kind of turn this is: answer, plan, or work.
  • Permission mode decides how much runs without asking you once a turn is working.

A turn in Agents mode under Default permission still stops at the first write or command to ask. A turn in Agents mode under Take the Wheel does not stop at all. See permission modes for what each of Default, Auto Review, Take the Wheel and Plan allows.

Where modes apply

The three chat modes are part of Perch Terminal, which is the desktop app and the CLI. Perch AI Web has no mode menu: it answers and, with web search on, searches, and it does not have local files or a terminal to act on.