Mirror CLI is Reflection AI's coding agent for the terminal. You describe a task and the agent reads your code, edits files and runs commands in the project until the task is done, showing each step. Reflection open-sourced it on October 16, 2026 alongside Beam 1.0, its first open-weight model: a 501B-parameter sparse Mixture-of-Experts model with 23B active parameters, released on Hugging Face under the Apache 2.0 license in BF16, FP8 and NVFP4 formats. Reflection says Beam was trained along with the Mirror harness. Mirror is labelled early beta, and commands and behaviour may change between releases.
By default Mirror runs in manual mode and pauses before shell commands, file writes and edits, connected-service tools and most extension tools; auto mode lets the agent proceed. /worktree starts another agent in a separate checkout so parallel tasks do not collide, and each workspace keeps its own conversation, model, reasoning effort and approval mode. /sandbox runs file and shell tools in a microsandbox microVM that reaches only allowlisted HTTPS hosts and gets no host credentials. Conversations are saved under ~/.mirror and can be resumed or branched, and project instructions (AGENTS.md), skills, MCP connections and Python extensions customise the agent.
Mirror uses Reflection's Beam-1.0-501B-A23B by default, but the provider is selectable: Reflection, OpenRouter, OpenAI Responses, Anthropic Messages, an existing ChatGPT plan through a Codex CLI login, or any compatible Chat Completions endpoint, including a local or self-hosted server. Architecturally, a Python product layer talks to a Rust agent core over a local connection.
Mirror itself is free and installs with uv tool install reflection-mirror on macOS 15+ (Apple Silicon) or Linux (x86-64/ARM64, glibc 2.28+); native Windows is not supported. Model usage is billed by the chosen provider; because Beam's weights are open, you can also point Mirror at your own Beam endpoint. Reflection describes its first-party API as a public beta that is free to start and does not publish per-token prices; usage is governed by rate limits and a plan-based daily token allowance.
