FigMirror

Turn your data into an editable matplotlib figure and PDF styled after a reference paper.

Stars
★ 574
Last updated
1mo ago
Primary language
Python

At a glance

How it runs
Agent plugin / skillWeb appCLI
Works with
Portable with changesCodex · Claude Code
Cost
Free software; you pay for model usage
Setup effort
Medium · a few setup steps
You'll need
Python 3uvCodex or Claude CodeShell / CLINetwork accessLocal filesystem
Typical use
A research author has a figure from a paper and wants to plot new experimental data with a closely related layout and visual style while retaining editable matplotlib code.
Not a fit if
  • Teams standardized on other OpenAI-compatible or local agent backends; only Codex and Claude Code are supported
  • Researchers needing reproducible benchmark artifacts; PlotTwin-Bench outputs are unreleased
  • Security-controlled environments that cannot review a piped remote install script
Website
arxiv.org/abs/2608.28814 ↗ set on GitHub
Maintenance
Commits in the last 6 monthslast commit 2026-09-30latest release v0.1.0 (2026-05-23)

What does this agent do, and when should you use it?

FigMirror reproduces the visual style of a reference paper figure using a user's data, producing an editable matplotlib script and a camera-ready PDF. It is delivered as skills for Codex and Claude Code, alongside a local browser interface for uploads, previews, iteration history, and refinement. A top-level Orchestrator runs a bounded loop between the figmirror-drawer and figmirror-reviewer agents; both harnesses share the same decision state machine and differ only in subagent transport. The Reviewer examines a far-view composite and full-resolution source and candidate images, then returns structured feedback, bounding boxes, a revision checklist, and a preserve list for the next Drawer pass. The workflow covers 2D and 3D figures and stores its artifacts in a local workspace. Its paper reports leading combined scores of 72.7 and 76.4 on a 150-instance PlotTwin-Bench evaluation subset, although the complete reproducibility artifacts remain unreleased.

The user supplies a screenshot of a paper figure and the data to plot. The Orchestrator delegates candidate generation to figmirror-drawer, which uses Grounded Measurement to identify visual coordinates and turn them into executable checks such as cropping a segment and reading pixel colors before producing and running matplotlib code. The figmirror-reviewer compares the reference and candidate at far and full-resolution views, returning a structured audit with bounding boxes, requested revisions, and elements that should be preserved. The Orchestrator converts that feedback into an annotated image and notes for the next Drawer pass, while the preserve list accumulates across iterations to resist style drift. The loop follows a shared decision state machine with a hard iteration cap and ultimately emits an editable matplotlib script and PDF. For 3D figures, the prompting and repair checks additionally address camera position, scale, surfaces, lighting, and geometry.

  1. A research author has a figure from a paper and wants to plot new experimental data with a closely related layout and visual style while retaining editable matplotlib code.
  2. A lab needs consistent typography, colors, composition, and visual hierarchy across result figures, with scripts that can remain part of a reproducible workflow.
  3. A user without a suitable reference can begin with the linked gallery of 139 paper figures spanning 25 chart families.
  4. A researcher recreating dense multi-panel layouts or joint hexbin plots with marginal histograms can use the Drawer–Reviewer loop to generate and audit successive candidates.
  5. An author preparing a 3D waterfall or similar scientific plot can use the geometry-aware checks for camera, scale, surfaces, and lighting.
  6. A Codex or Claude Code user who prefers a browser workflow can run the local Web UI to upload inputs, compare iterations, and request refinements.

Key setup facts

Pulled from the install and usage notes below. The project README has the current values.

Install / run
  • curl -fsSL https://raw.githubusercontent.com/VILA-Lab/FigMirror/main/scripts/install.sh | bash
Ports
8765
Environment variables
UV_CACHE_DIR

How do you install or deploy this agent?

For the complete local Web UI:

git clone https://github.com/VILA-Lab/FigMirror.git && cd FigMirror
bash scripts/install.sh
uv run python scripts/figcopy_serve.py --workspace .artifacts/figmirror-workspace --backend codex

Then open http://127.0.0.1:8765/. Replace the final option with --backend claude to use Claude Code. If uv is unavailable, first run python3 -m pip install uv.

For a skill-only installation:

curl -fsSL https://raw.githubusercontent.com/VILA-Lab/FigMirror/main/scripts/install.sh | bash

To select a runtime explicitly, pipe the same URL to bash -s -- --codex, bash -s -- --claude, or bash -s -- --all. The installer detects Codex and Claude Code and installs the figmirror skill plus the figmirror-drawer and figmirror-reviewer custom agents. On shared machines or systems with a small root filesystem, run df -h and point UV_CACHE_DIR to a large user-writable filesystem before the first invocation. No standalone API credential setup is documented; an operational Codex or Claude Code environment is required.

How do you use this agent?

In skill mode, attach a screenshot of the reference paper figure, paste the data, and ask Codex or Claude Code: Use FigMirror to mirror this figure's style with my data. FigMirror then runs the Drawer–Reviewer loop and produces editable matplotlib code and a PDF. For the Web UI, start scripts/figcopy_serve.py, visit http://127.0.0.1:8765/, and use the browser to upload inputs, preview candidates, inspect iteration history, and refine the result. The documented workspace example is .artifacts/figmirror-workspace.

What are this agent's strengths and limitations?

Pros
  • It delivers both editable matplotlib source and a PDF, rather than limiting the output to a flattened image.
  • The role-separated Drawer and Reviewer exchange bounding boxes, revision tasks, and a persistent preserve list, giving each iteration explicit visual feedback.
  • Codex and Claude Code share one bounded decision process and can be selected as backends for the same local Web UI.
  • Grounded Measurement converts visual coordinate estimates into executable operations such as crops and pixel-color checks instead of relying only on prose judgments.
  • The documented scope includes complex 2D compositions and 3D figures, with dedicated checks for camera, scale, surfaces, and lighting.
  • A gallery of 139 figures across 25 chart families provides starting references for users who do not already have one.
Limitations
  • The only documented production harnesses are Codex and Claude Code; support for additional OpenAI-compatible or local-agent backends remains a roadmap item and would require adaptation.
  • Deployment requires Python, uv, shell access, filesystem access, and a working Codex or Claude Code installation; small-root or shared machines may need custom cache placement.
  • The paper reports PlotTwin-Bench results, but the complete references, target data, generated outputs, and evaluation metadata needed for reproduction have not yet been released.
  • The repository license is unknown, so organizations must resolve usage, modification, and redistribution rights before adoption.
  • The multi-pass generation and review process has a hard cap, but the supplied material does not quantify its runtime cost, latency, or failure rate.
  • The quick installer either downloads the repository or pipes a remote script into a shell, which may require prior review in security-controlled environments.

How does this agent compare with similar options?

The paper compares FigMirror with Plot2Code, METAL, ChartGalaxy, and ChartIR in quantitative and qualitative evaluations. All methods in the reported comparison use GPT-5.5; FigMirror records the highest combined score on both splits, scoring 72.7 on 50 hand-curated references and 76.4 on 100 randomly sampled augmented references. The supplied material does not describe the alternatives' deployment models or feature sets, so it does not support a broader comparison of adoption cost or compatibility.

Key facts side by side with the most closely related agents.

Agent Source review Form / cost Stars Updated Language Full support on
FigMirror This agent 46 · Major gaps Agent plugin / skillFree + model costs ★ 574 1mo ago Python Codex · Claude Code
Strategy Visualization Workbench 74 · Some gaps Agent plugin / skillFree + model costs ★ 177 1mo ago Python Codex · Claude Code
Auto-Research-Skills Research Toolkit 61 · Some gaps CLIFree ★ 265 today Python —
Feynman Research 53 · Major gaps CLIFree + model costs ★ 10k 5d ago TypeScript —

How does FollowAgents rate this agent?

FollowAgents source review · FARS-2.1
Major gaps
46/ 100 5-point scale 2.3 / 5
Trust 11/29
Reliability 5/14
Adaptability 12/18
Convention 8/18
Effectiveness 7/13
Verifiability 3/8
Why each dimension lost points
Trust11 / 29 · 1.9/5

The materials describe a local Web UI, a workspace, and uploaded reference images; they also require echoing the parsed data and normally asking for confirmation before drawing, supporting user control over input data and earning 2. The install instructions also pipe a remote script into a shell and the workflow writes artifacts, while the packet does not show permission boundaries, data retention or transfer details, dependency review, external-effect disclosures, rollback, or clear attribution and ownership; the other trust criteria therefore receive 1.

Reliability5 / 14 · 1.8/5

The materials specify iteration caps, role separation, and structured review, but the current and retired skill files describe different architectures, and a test configuration comment says dependencies are empty despite runtime dependencies in pyproject. The packet does not establish dependency availability or the quality of failure messages, so each criterion receives 1.

Adaptability12 / 18 · 3.3/5

The intended use and required inputs are clear, 2D and 3D style transfer are covered with a 3D gate, and generic plotting without a reference is explicitly out of scope. Both runtimes and a local Web UI have install instructions, supporting scores of 2. The packet gives less detail about user roles and environment compatibility, so higher scores are not justified.

Convention8 / 18 · 2.2/5

The README, install guide, and skill instructions provide navigable structure, commands, an example request, and basic setup troubleshooting. However, the packet includes both current and retired skill material, documents few limitations, provides only partial version history, does not clearly identify maintenance ownership, and supplies neither license metadata nor a license text, so these criteria are deducted.

Effectiveness7 / 13 · 2.7/5

The stated outputs include an editable matplotlib script, PNG, and PDF, with iteration browsing, refinement, and showcase examples. Reference-conditioned style transfer has a clear added value, but the costs and resource requirements of the agent and image/model workflow are not adequately described, so cost-benefit receives 1.

Verifiability3 / 8 · 1.9/5

The README reports paper and benchmark results and links to the corresponding materials, but the packet does not include the paper contents or the full implementation needed to cross-check those results. Claims are not consistently separated from demonstrated facts, so traceability, cross-source corroboration, and fact/inference separation each receive 1.

Risks and how to mitigate them
  • The install command downloads a remote script and pipes it to a shell; the reviewed materials do not describe the script's permission scope or how uploaded references and data are handled.
  • The README, current skill, and retired skill materials differ in their architecture descriptions. This assessment is based only on the supplied static packet and is not runtime verification.
  • License metadata is unknown and no license text is included in the packet; verify licensing before use or redistribution.
Evidence confidence: Low Reviewed Nov 08, 2026 Reviewed revision 11631cc31a3f
See the full review method →
Maintain this project? Add the FARS badge to your README

The badge shows the current review score and verdict and links back to this page. It refreshes about once an hour after a re-review.

FARS

[![FARS review](https://followagents.com/badge/figmirror.svg)](https://followagents.com/en/agents/figmirror)
<a href="https://followagents.com/en/agents/figmirror"><img src="https://followagents.com/badge/figmirror.svg" alt="FARS review"></a>

Recent releases

Latest entries from the project's GitHub Releases. Notes are quoted from upstream as written, not rewritten or tested by FollowAgents.

  1. v0.1.0 FigMirror v0.1.0: Public Preview
    Release notes excerpt
    FigMirror is now publicly available. FigMirror helps you plot your own data in the visual style of a reference paper figure. Pick a reference figure, provide your data, and get an editable matplotlib script plus a camera-ready PDF. Highlights: - Reference-to-figure Drawer / Reviewer loop - Local Web UI for uploading data, browsing iterations, and refinement - Codex and Claude Code skill support - 139-figure gallery for starting from real paper-style examples - English and Chinese README - Contribution guide for showcase cases, prompt improvements, UI polish, and evaluation cases …
    View on GitHub ↗

FAQ

Do I have to use Codex or Claude Code?
Those are the only documented production paths. The Web UI likewise selects either --backend codex or --backend claude; broader OpenAI-compatible and local-agent backend support is still on the roadmap.
Does it require an API key or paid model access?
The supplied material does not document API-key configuration, pricing, or a free execution path. It requires an operational Codex or Claude Code environment, so authentication and cost depend on that environment.
Can I edit the generated result?
Yes. The stated outputs include an editable matplotlib script as well as a camera-ready PDF.
Will it iterate indefinitely until the figure is an exact match?
No. Both harnesses use a decision state machine with a hard iteration cap. The cap's numeric value is not given, and pixel-perfect equivalence is not promised.
Can I reproduce the published benchmark results today?
Not from the supplied materials alone. The paper reports results on a 150-instance subset, but releasing the complete PlotTwin-Bench references, target data, outputs, and evaluation metadata is still a roadmap item.
Is FigMirror open source?
The source code is public on GitHub (VILA-Lab/FigMirror), but the repository does not declare a recognised licence, so the rights to reuse or modify it are unclear. Ask the author or check the README before building on it.
Is FigMirror safe to use?
FollowAgents' FARS review (Nov 08, 2026) scored FigMirror 11/29 on trust and safety. None of the eight safety controls the review checks for was missing outright.
Is FigMirror still maintained?
Development has slowed: the latest commit on GitHub is from 2026-09-30, within the last six months but not the last 30 days. The latest release is v0.1.0 (2026-05-23).
Who makes FigMirror?
FigMirror is published on GitHub by the account VILA-Lab. FollowAgents has not verified which person or company is behind that account.
What is FigMirror's official website?
The project's GitHub repository lists https://arxiv.org/abs/2608.28814 as its website. That link is set by the repository owner; the source code itself is on GitHub (VILA-Lab/FigMirror).
Was this page helpful?
View on GitHub ↗ Install ↓

Compare agents like this one

The same FARS review applied across the shortlist this agent qualifies for.

Related concepts