Skip to content
Back to search

Wiggle

Wiggle is a sandbox to run code in

Safety scan

A static text scan, not a guarantee. Select the badge to see what was scanned and every finding.

Checked today
Badge for your README
findsafeskills safety rating

Markdown

[![findsafeskills safety rating](/api/listings/razshy--wiggle/badge.svg)](/listings/razshy--wiggle)

HTML

<a href="/listings/razshy--wiggle"><img src="/api/listings/razshy--wiggle/badge.svg" alt="findsafeskills safety rating" /></a>

Install

i

Install instructions from the README

As written by the repo's author. findsafeskills has not verified these steps.

docker build -t wiggle .
docker run -it wiggle bash

# inside:
soffice --version            # LibreOffice 24.2
python3 -c "import pandas"   # full data science stack
ls /mnt/skills               # the 40 playbooks Claude follows

Want the exact bytes instead of a rebuild? The [GitHub release](https://github.com/Razshy/Wiggle/releases/tag/v1.0-live-rootfs) carries the full filesystem of a live session (8.9 GB, sanitized):

cat wiggle-part-aa wiggle-part-ab > wiggle.tar.zst
zstd -d -c wiggle.tar.zst | docker import - wiggle:live

See [ROOTFS.md](ROOTFS.md) for the full download and verification steps.

What was scanned

The files the safety scan read, as written by the repository's author and shown as plain text. Lines with a finding are marked.

README.md · README · 139 lines

View this file on GitHub (read at the repository's current version: the scan did not record which commit it read)

  1. # Wiggle
  2. A working copy of the sandbox that Claude (claude.ai) runs code in.
  3. When Claude executes a command, reads a file, converts a document, or takes a
  4. screenshot, it is not running on a web server. It is running inside a small
  5. virtual machine, one per conversation. This repository is that machine: the
  6. same files, tools, packages, and instructions, packaged so you can run it
  7. yourself on Docker, e2b, Modal, or any Linux host.
  8. Codename `wiggle` is what Anthropic calls this VM internally.
  9. ## Quick start
  10. ```bash
  11. docker build -t wiggle .
  12. docker run -it wiggle bash
  13. # inside:
  14. soffice --version # LibreOffice 24.2
  15. python3 -c "import pandas" # full data science stack
  16. ls /mnt/skills # the 40 playbooks Claude follows
  17. ```
  18. Want the exact bytes instead of a rebuild? The
  19. [GitHub release](https://github.com/Razshy/Wiggle/releases/tag/v1.0-live-rootfs)
  20. carries the full filesystem of a live session (8.9 GB, sanitized):
  21. ```bash
  22. cat wiggle-part-aa wiggle-part-ab > wiggle.tar.zst
  23. zstd -d -c wiggle.tar.zst | docker import - wiggle:live
  24. ```
  25. See [ROOTFS.md](ROOTFS.md) for the full download and verification steps.
  26. ## How it works
  27. - Claude decides to run something and sends the command over a private
  28. channel to a supervisor process inside the VM.
  29. - The supervisor starts the command with a fixed environment (see
  30. [meta/env-contract.md](meta/env-contract.md)) and streams the output back.
  31. - The command runs as root with no seccomp and no user sandbox. Inside this
  32. VM, root is normal. The isolation that matters is the VM boundary itself.
  33. - User files arrive as mounted folders under `/mnt/user-data`. In production
  34. those are remote storage; here they are plain directories.
  35. Machine spec to match if you care about parity: 1 vCPU, 3.9 GiB RAM, no swap.
  36. Details in [meta/box-spec.md](meta/box-spec.md).
  37. ## What is inside
  38. | Capability | What gives you that |
  39. |---|---|
  40. | Documents (docx, xlsx, pptx, pdf) | LibreOffice 24.2, pandoc, python-docx, openpyxl |
  41. | OCR (make scans searchable) | tesseract 5 with English models, byte-identical to the public Google release |
  42. | File-type detection | magika (Google model, public build) |
  43. | Web automation and screenshots | Playwright 1.56 with a pinned Chromium at `/opt/pw-browsers`, plus puppeteer with Chrome-headless-shell |
  44. | Diagrams and charts | mermaid-cli, matplotlib, graphviz |
  45. | Numbers and data | pandas, numpy, scipy, scikit-learn, Jupyter |
  46. | Typesetting and PDFs | TeX Live 2023, poppler, qpdf, ImageMagick, wkhtmltopdf |
  47. | Media | ffmpeg |
  48. | **The playbooks Claude follows** | [mnt-skills/](mnt-skills/), 40 skills, verbatim |
  49. The skills are the interesting part. They are plain Markdown instruction
  50. files that Claude reads before doing certain jobs: how to fill a PDF form,
  51. how to run deep research (including the sub-agent prompts), how to drive the
  52. desktop with computer use, how to build a skill. Public ones cover office
  53. documents and file reading; example ones cover deep-research, morning
  54. briefings, painting, MCP server building, and more. Browse them under
  55. [mnt-skills/](mnt-skills/).
  56. Two custom Anthropic binaries are included: `extract-text` (turns uploaded
  57. documents into text, Rust) and the mount daemon interface (`rclone-filestore`,
  58. Go, documented in [meta/filestore-api.md](meta/filestore-api.md)).
  59. ## What is not included
  60. Four things live outside the filesystem, so no dump could contain them:
  61. - The supervisor binary itself (runs from RAM, never from disk).
  62. - The model, which is remote by definition.
  63. - Anthropic's egress firewall and its CA roots.
  64. - The remote storage service behind `/mnt/user-data`, whose contract is
  65. documented in [meta/filestore-api.md](meta/filestore-api.md).
  66. For e2b users: e2b's own daemon takes the supervisor's role, which is why
  67. this image drops straight into an e2b template.
  68. ## Using it with your own agent
  69. Any agent that can shell into a container can use this box exactly the way
  70. Claude does: read the relevant `/mnt/skills/*/SKILL.md`, then run the tools it
  71. names. Inject the environment from
  72. [meta/env-contract.md](meta/env-contract.md) and the behavior matches
  73. production, quirks included.
  74. ## Known quirks (present in the original, kept on purpose)
  75. - `pandoc x.md -o x.pdf` fails until you add `fonts-lmodern` or use the
  76. Chromium fallback. The real sandbox has the same hole.
  77. - `pip install` is blocked by PEP 668. Use `uv`. The offline wheel cache is
  78. included.
  79. - ImageMagick has no SVG coder and ignores `-quality` for webp. Use Pillow or
  80. sharp.
  81. - Tool exit codes lie often. Verify outputs (`test -s out && file out`)
  82. instead of trusting success.
  83. - `extract-text` segfaults under x86 emulation on ARM Macs. It is fine on
  84. native x86-64 hosts.
  85. ## Repository layout
  86. - [Dockerfile](Dockerfile) the recovered build recipe (verified to build green)
  87. - [_context/](_context/) build inputs: pinned manifests and the Anthropic binaries
  88. - [mnt-skills/](mnt-skills/) the 40 skills, copied into the image at `/mnt/skills`
  89. - [meta/](meta/)
  90. [dpkg-manifest](meta/dpkg-manifest.txt),
  91. [uv-manifest](meta/uv-manifest.txt),
  92. [env-contract](meta/env-contract.md),
  93. [filestore-api](meta/filestore-api.md),
  94. [box-spec](meta/box-spec.md)
  95. - [ROOTFS.md](ROOTFS.md) how to get the exact live filesystem dump
  96. - [LICENSE](LICENSE) and [NOTICE](NOTICE)
  97. ## License
  98. The build recipe, scripts, and documentation in this repo are MIT
  99. ([LICENSE](LICENSE)).
  100. Everything Anthropic-made is theirs, not MIT:
  101. - The 40 skill files in [mnt-skills/](mnt-skills/) and the Anthropic binaries
  102. each carry Anthropic's own license ([NOTICE](NOTICE) and each
  103. `LICENSE.txt`), which does not clearly allow redistribution. They are here
  104. as captured research artifacts.
  105. - If you would rather not ship them, delete `mnt-skills/` and remove the one
  106. `COPY mnt-skills/` line from the Dockerfile. The box still builds and runs;
  107. any user can re-obtain the skills by asking Claude to show them.
  108. Not affiliated with Anthropic.

Ratings and comments

Sign in to rate this skill.

Sign in

Free to use. If it’s helped, pass it on.