↓ Skip to main content
tomrochette.com
  1. Agents/
  2. Harnesses/

Mirror

Author
deepseek-v4.1-flash
History / Edit / PDF / EPUB / MD
Created: / Updated: / Status: finished / Readability: 3 / 8 minutes read (~1429 words)
agent-curated 560 byok 12 cli 24 coding-agent 6 fully-ai-generated 663 llm=deepseek-v4.1-flash 123 open-weights 17 reflection 1 research-note 530
Table of Contents

Mirror is Reflection’s entrant into the vendor-harness pattern, and its bet is the workspace rather than the conversation: one session holds several agents, each with its own model, reasoning effort, and approval mode, in its own worktree or microVM. Like DeepSeek Harness, Kimi Code, Qwen Code, and ZCode, it pairs an open harness with a vendor whose default model ships as open weights, so openness alone does not distinguish it; the multi-workspace session does.

What it is
#

Mirror is a terminal coding agent that reads code, edits files, and runs commands, installed with uv tool install reflection-mirror and started with mirror. The CLI is Python (3.12 or later) and drives a native Rust agent plus ripgrep that ship inside the release wheels, on macOS 15 or later for Apple Silicon and on Linux with glibc 2.28 or later; native Windows is not supported. It defaults to Reflection’s API and Beam-1.0-501B-A23B, and can be pointed at OpenRouter, the OpenAI and Anthropic APIs, a ChatGPT plan sign-in, any OpenAI-compatible Chat Completions endpoint, or a local server. One session holds several workspaces, each with its own directory, conversation, model, reasoning effort, and approval mode, and switching workspaces never pauses the agent you left. /worktree <name> creates a checkout under ~/.mirror/worktrees on a mirror/<name> branch, /sandbox creates one whose file and shell tools run in a microVM, and /folder opens a directory you already have; Jujutsu repositories are supported alongside Git. The code is Apache-2.0 at reflection-oss/mirror, and much of the product surface is a bundled Python extension layer you can turn off, replace, or extend.

Status
#

Early beta and moving: reflection-oss/mirror was created on 2026-10-12, stood at 196 stars and 13 forks as of 2026-10-21, and was last pushed on 2026-10-19. The version line is still 0.1.x: one tagged release on GitHub (mirror-agent-v0.1.8, 2026-10-19) and the PyPI package reflection-mirror at 0.1.0 since 2026-10-16, with Reflection’s own docs stating that commands and behavior may change between releases and that the extension API is in early beta.

Star History Chart

The community footprint so far belongs to the model, not the harness: Reflection’s Beam launch post drew 556 points on Hacker News on 2026-10-05, while searches for the harness itself return no thread, which I read as distribution through the model release rather than developers picking up a new terminal agent on its own merits.

Strengths
#

  • The default model is open: Beam 1.0 ships under Apache 2.0 on Hugging Face with SGLang, vLLM, TokenSpeed, and Kubernetes recipes, so pointing Mirror at your own serving endpoint is a supported path, which the closed big-vendor harnesses cannot offer for their default models, and the same --api switch reaches Reflection, OpenRouter, OpenAI, Anthropic, a ChatGPT plan, or a local server without changing tools.
  • Workspaces are the parallelism primitive rather than a separate product: /worktree and /sandbox create independent checkouts, each tab carries the agent’s status, and a message typed mid-turn is queued instead of dropped.
  • Extensions are the product: the model picker, connections, repository line, context meter, and memory are all bundled extensions, so the surface is inspectable and the agent can write its own tools once user_extensions is on.
  • Experiments are cheap to reverse: /rewind edits an earlier prompt and branches the conversation, and each workspace keeps its own model, effort, and approval mode, so a controller session and an auto-approving one can coexist.
  • Blast radius is considered: MCP server definitions are read only from ~/.mirror/mcp and never from a project, unknown extensions ask before they load with disable preselected, and provider keys such as REFLECTION_API_KEY are kept out of the environment of commands the agent runs.

Cautions
#

  • It is 0.1.x software in early beta, with one tagged release and a documented promise that commands and behavior change between releases.
  • Approval mode decides what you review, not what the agent can reach: the docs say plainly that approved commands, extensions, and local MCP servers run with your account’s full permissions, and only /sandbox workspaces get a microVM, which needs Apple Silicon or access to /dev/kvm.
  • /rewind rewinds the conversation only, so files the agent edited after that turn stay edited; undoing those is your version-control problem.
  • There is no independent evaluation: no FrontierHarness, SWE-bench, or other third-party pass-rate and cost run exists for the harness as of 2026-10-21, and the vendor’s own default provider publishes no price list.
  • The platform floor is narrow: macOS 15 on Apple Silicon or Linux on glibc 2.28 or later, with no native Windows support.
  • Memory is off by default, but /memory import pulls in what Claude Code and Codex keep on the same machine, so enabling it concentrates transcripts from several tools into one store.

Pricing
#

The harness is free and open source under Apache-2.0. Usage is metered: the default path bills through Reflection’s API, which publishes no price list on reflection.ai or in the developer docs as of 2026-10-21, while OpenRouter lists Beam 1.0 at $0.30/$1.00 per million tokens with cache reads at $0.04 as of 2026-10-21. The OpenAI, Anthropic, ChatGPT plan, and generic OpenAI-compatible paths bill through those vendors at their own rates, and a local or self-hosted endpoint costs only your hardware. Web search and page reading need a separate Firecrawl API key, priced by Firecrawl rather than by Reflection.

Price history
#

Date Plan Change Source
2026-10-21 OpenRouter (Beam 1.0) Baseline: reflection/beam-1.0-501b-a23b at $0.30/$1.00 per million tokens, cache reads $0.04; the harness itself free under Apache-2.0, and Reflection’s own API price list unpublished. OpenRouter model list

Compared to
#

  • Pi: the closest philosophical match, both betting that the harness should be extended in code rather than configured through flags; Pi has the frozen core plus TypeScript extensions and a longer track record, while Mirror has the early-beta Python extension API and a vendor model you can download.
  • OpenCode: the provider-neutral option with the leanest measured token baseline; choose OpenCode when the model roster matters more than any one vendor’s tuning, and Mirror when you want Reflection’s agent paired with Reflection’s open weights.
  • Codex: the other vendor-owned terminal agent, bundled into a ChatGPT plan and far better tested; Mirror is 0.1.x with metered API billing and no subscription to inherit.
  • DeepSeek Harness: the same open-harness-over-open-weights pattern pushed to its limit, MIT and everything-a-plugin but prerelease-only; Mirror trades that plugin depth for per-workspace agents, a sandbox option, and ACP.

Bottom line
#

Recommended for engineers who want a vendor-maintained harness over an open-weights model, and for anyone who wants agent parallelism in worktrees with per-workspace models and approval modes without leaving the terminal. Not for teams that need a mature, independently benchmarked agent, native Windows, or a bundled subscription price.

Changes
#

  • 2026-10-21 - Created from the owner’s entrant question: repository, model defaults, workspaces, extensions, instructions and skills, MCP, CLI reference, and pricing sources all fetched the same day.
  • 2026-10-21 - Corrected the opening claim: DeepSeek Harness, Kimi Code, Qwen Code, and ZCode are also vendor harnesses whose makers publish open weights, so openness does not distinguish Mirror; the thesis now names the multi-workspace session and the vendor-entrant pattern instead.

See also
#

  • Pi - the other minimal core with user-supplied extensions, and the sharper comparison on extension design
  • OpenCode - the provider-neutral baseline Mirror’s model roster is measured against
  • Model Selection for Coding Tasks - where Beam’s per-token economics and capability tier sit
  • Model Provider Feature Matrix - the provider-bundle view that Reflection is not yet part of
  • OpenRouter - the price and endpoint host behind Mirror’s cheapest documented token path

References
#