Reflection open-sources Mirror, its terminal coding agent for parallel worktrees
Reflection AI has released Mirror, a Rust-based terminal coding agent that reads and edits files, runs commands and can work on parallel tasks across separate repository worktrees, under an Apache-2.0 licence.
Image: GitHub
Why it mattersA coding agent that runs several tasks on separate repository worktrees at once lets a developer give the agent multiple branches to try without the tasks overwriting each other's files.
Reflection AI, the maker of the Beam open-weight model released earlier this month, published its own terminal coding agent on 12 October 2026. The repository, at reflection-oss/mirror on GitHub, is Apache-2.0 licensed and had reached 176 stars by six days after launch.
The README positions Mirror as a coding agent for the terminal that can read and edit files, run commands and work on parallel tasks in separate repository worktrees. The native agent itself is written in Rust, and the Python package reflection-mirror ships release wheels that bundle the agent and ripgrep so no local Rust toolchain is required.
What is in the release
Mirror is installed with the uv package manager as uv tool install reflection-mirror. The command on the shell is mirror. The README lists supported platforms in the CLI subdirectory and tells users to run mirror inside the project directory they want the agent to work in.
A separate mirror-dev script runs the agent from a source checkout for developers who want to build and iterate on the agent code itself. That path needs uv, Rust and Cargo, a C compiler, protoc, and ripgrep on the system path. The README notes that the Rust checkpoint tests also require Jujutsu, which the project calls out as jj on PATH.
The parallel worktree model
The feature that distinguishes Mirror from the growing list of terminal coding agents is working on multiple tasks at once in separate repository worktrees. A worktree is a git feature that gives a repository several checked-out branches on disk at the same time, each with its own working directory, so files on one branch cannot overwrite files on another.
The maintainers do not quote a benchmark or a measurement for the design in the README. The documentation link in the file points at developers.reflection.ai/mirror-overview for a longer explanation of how parallel tasks are managed.
Permissions, and the model vendor pattern
The README states that by default tools run with the user's own account permissions, and that manual approval is not a sandbox. There is a link to a local data and safety section in the CLI README for teams that want to restrict what the agent can touch. The license is Apache-2.0, with the preserved-notices clause the project describes.
Mirror arrives from the same company that released the 501B-parameter open-weight model Beam under Apache-2.0 two weeks before this coding agent. That gives a team using Beam a first-party agent to call it from, matching the pattern Anthropic and OpenAI follow with Claude Code and Codex, where the model vendor also publishes the agent that uses it.
Source
reflection-oss/mirror on GitHub. Documentation at developers.reflection.ai/mirror-overview.
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.