Skip to main content
Mirror CLI is Reflection’s coding agent for the terminal. You describe a task, and the agent reads your code, edits files, and runs commands in your project until the task is done, showing each step as it goes. It uses Reflection and Beam-1.0-501B-A23B by default. View on GitHub To install Mirror and connect it to Reflection, follow the quickstart. This page explains how Mirror works once it’s running.
Mirror CLI is in early beta, and commands and behavior may change between releases. Run /help to see the commands in your installation.

How Mirror is organized

Run mirror in a project directory to start a Mirror session. The first time you start Mirror in a directory, it asks you to confirm that you trust that directory’s files. One Mirror session holds all of the following:
  • Workspace. A directory with its own agent and its own conversation. Mirror starts with one workspace for the directory you launched from. You can add more with /worktree, which creates a separate checkout so that two agents can edit files without getting in each other’s way. See workspaces.
  • Conversation. The history of prompts, agent replies, and tool calls in a workspace. Mirror saves every conversation in ~/.mirror, so you can resume one later or branch from an earlier turn.
  • Tools. How the agent acts on your project: reading and searching files, editing them, and running shell commands. Tools run on your machine with your account’s permissions. Connections to services such as Linear or Notion, and extensions, can add more.
  • Instructions and skills. Guidance the agent reads before it starts, from AGENTS.md files and reusable skill directories. See instructions and skills.
Workspaces are independent: switching to another one doesn’t pause the agent you left, and each keeps its own model, reasoning effort, and approval mode.

The Mirror screen

Mirror v0.1.0 welcome screen showing the model, working directory, prompt, settings line, and clocks

The Mirror welcome screen with the model, working directory, and prompt.

From top to bottom:
  1. Transcript. The conversation so far. Before the first prompt it shows a welcome screen with the model and working directory. Use Ctrl+T or /view to switch between detail (full tool output and reasoning), default (grouped tools and reasoning), and focus (hide tool activity and reasoning).
  2. Prompt. Where you type. Extensions can show widgets above or below it.
  3. Settings line. The approval mode (manual or auto), the model, and the reasoning effort on the left, and generation speed and context use on the right.
  4. Repository line. The current branch and a count of uncommitted changes.
  5. Workspace tabs. Shown once you have more than one workspace, with a status indicator for each agent.

Work with the agent

Choose what to approve

By default Mirror runs in manual mode: it pauses before shell commands (including ones that only read, such as ls), file writes and edits, connected-service tools, and most tools that extensions add. It shows the call, and you press 1 or Enter to approve, or 2 to reject. Reading and searching files with the agent’s built-in tools never needs approval. Pressing Escape at an approval prompt stops the whole turn, not just that call. Switch to auto mode to let the agent use tools without routine approval prompts. Tools that require explicit confirmation still ask, and external extensions need consent before loading. To switch, do any one of these:
  1. Press F2 to toggle between manual and auto.
  2. Run /approvals to cycle between manual and auto.
  3. Start Mirror with mirror --approval-mode auto.
Mirror remembers your choice and uses it the next time you start Mirror, in any directory.
Approval mode decides what you review. It isn’t a sandbox. Approved commands, extensions, and local MCP servers run with your account’s full permissions.
Mirror does keep its own provider API keys, such as REFLECTION_API_KEY, out of the environment of commands the agent runs.

Conversations

Each workspace has one active conversation. Mirror saves it as you go, so nothing is lost when you exit.
Rewinding changes only the conversation. Files the agent edited after that point stay as they are. Use version control to undo file changes.

Next steps

Workspaces

Run several tasks in parallel, each in its own checkout.

Models and providers

Change the model, reasoning effort, or provider.

Instructions and skills

Give the agent project guidance and reusable procedures.

Connections

Connect services such as Linear and Notion, or your own MCP servers.

Extensions

Turn bundled features on and off, or write your own in Python.

Reference

Look up commands, shortcuts, launch options, and files.
Last modified on November 3, 2026