abox.tools for agents

Published by Abox Technologies. Compress, resize and convert images; create ICO icons, rasterize SVGs and compose image layouts; compress, watermark, merge, split, reorder and rotate PDF pages; extract PDF tables to CSV; turn images into PDFs; format JSON and XML; convert YAML and JSON; compare text files; encode or decode Base64; embed images as data URIs; generate QR codes and barcodes; inspect or remove image metadata; and calculate file checksums with the existing abox.tools code. An agent supplies local paths and receives output paths and short JSON results. Offline file processing keeps bytes on your computer and out of tool responses.

It also processes GIFs, admitted audio and supported video timelines; straightens document photos; prepares custom ID photos and local previews; extracts receipt fields with bundled OCR; inspects and renders a restricted DICOM subset; protects and unlocks PDFs; redacts explicit image/PDF regions; and generates passwords into local files. Codec and rendering operations expose documented subsets.

The public v0.11.0 release provides 66 offline file operations, six opt-in live controls and one capabilities tool: 72 operations and 73 MCP tools. Native and universal packages passed checks on all four target platforms. See tool coverage and concrete unsupported paths for the source engines reused by this adapter and the limits of each route.

Version 0.11.0 adds six live-session controls: a separate opt-in visible companion for text/files, camera video and voice. Opening a window starts no network or capture. The person clicks Start/Connect and handles permissions and pairing there. Live mode uses the existing abox.tools pairing service and direct encrypted WebRTC; the service sees negotiation metadata and keeps seven-day Cloudflare metadata logs. There is no relay, discovery or recording. Read the live-session boundary. The actual SHA-verified public Windows cache passed all 72 operations/73 tools and all eight passive companion roles, with no capability absences or unrequested pairing/capture. Public Gemini installation and the 15-file shared-skill copy also passed. Real pairing-service and public-website interoperability checks used temporary files and generated media in two isolated browsers on one computer. Physical devices, native camera permission and desktop-client installation remain unverified. See v0.11.0 publication evidence.

Historically, version 0.10.0 added HEIC/AVIF, flattened PDF redaction, sound extraction and admitted H.264 video processing. It also expands audio and DICOM codec support. Publication and the Registry readback were verified on 28 October 2026. The default Python installer downloaded and checksum-verified the public Windows package. That actual cache passed all 66 file operations and 67 MCP tools with its bundled runtime, without capability absences or a local rebuild. Public Gemini CLI installation and the shared skill's 14-file project installation also passed. The skill check validated the installer contract offline; actual adapter installation was proved separately by the public Python download and full MCP smoke. See the v0.10.0 publication evidence.

Historically, the public v0.9.0 Windows download passed all 55 file operations and 56 MCP tools using its bundled runtime. Public Gemini CLI installation and the shared skill's ten-file installation were also checked. See publication evidence, including the historical v0.8.0 receipts.

Download the latest release. Listed in the Official MCP Registry as io.github.A-Box-of-Tools/abox-tools. New tagged releases update the Registry automatically after their downloads pass verification.

Release bundles include Node.js, production dependencies, pinned processing code, and the shared skill. You do not need to install Node or npm, or type setup commands. Image rendering/conversion, GIF creation/extraction, decoded audio, receipt OCR, document scans, local previews, PDF raster redaction and video decoding/encoding need an installed Edge, Chrome or Chromium. PDF page/security operations, metadata operations, checksums, container inspection, audio/video packet copying and text processing run without one. QR codes and barcodes default to SVG without a browser; PNG output uses the installed local browser. A native H.264 encoder is required for new MP4 encoding.

Install in Codex

Download the matching .zip from the release's Assets:

1. Extract the entire ZIP to a permanent folder. For example, a 0.11.0 Windows bundle extracts as abox-tools-0.11.0-windows-x64. Keep this folder in place because the local marketplace uses it for plugin refresh.

2. Open START-HERE.html inside the extracted folder. This page works offline. Select Install in Codex, allow your browser to open the desktop app, then select Install on the plugin's detail page.

3. Start a new local chat and ask for one of the operations below.

You need the desktop app installed with its codex:// link handler available. The button opens the app's supported local-plugin detail link; the app handles installation. See the official OpenAI deep-link documentation. If your browser blocks the link, use Install in Codex.cmd on Windows or Install in Codex.command on macOS. Linux includes Install in Codex.sh for file managers that support executing shell scripts. These launchers open the same detail page and may briefly show a terminal; no commands need to be typed.

Related MCP server: pdf-agent-toolkit-mcp

Install the tools in Claude Desktop

Download the universal.mcpb from the same release, open it with Claude Desktop, and follow its installation prompts. One file supports Windows x64, Linux x64, and both Apple silicon and Intel Macs. It includes the runtimes and processing dependencies. Smaller platform-specific MCPBs are also available. This installs the MCP tools; the shared agent skill is part of the Codex ZIP.

If opening the file does not show an installer, go to Settings → Desktop app → Manage beside Extensions, then Advanced settings → Install Extension… and select the .mcpb. See Claude's desktop extension instructions. Bundle format and MCP execution are checked in CI on all four platforms. The 0.2.1 universal bundle was installed and enabled in Claude Desktop on Windows, where the app reported its local MCP server Running. The macOS desktop interface has not been checked. Download and install a new bundle when updating; custom MCPB files do not update automatically.

Version 0.3.0 corrects the publisher to Abox Technologies. If your desktop host shows both the earlier publisher entry and the new one, disable the older entry to avoid duplicate tools.

Install in Gemini CLI

With Gemini CLI installed, run:

gemini extensions install https://github.com/A-Box-of-Tools/abox-tools-plugin

Gemini selects the matching prebuilt archive from the latest GitHub Release. The extension includes its runtime, dependencies and shared skill. Review the installation prompt and allow its local MCP server. See the Gemini extension release guide.

Install tools in Cursor

Extract the matching native ZIP, open START-HERE.html, and select Install tools in Cursor. Cursor shows the local server configuration before adding it. Keep the extracted folder in place. This adds the MCP tools; the shared skill can be installed separately. The repository's Cursor plugin route needs Node and a one-time adapter setup; see Cursor installation.

Install tools in VS Code

Extract the matching native ZIP, open START-HERE.html, and select Install tools in VS Code. Review the local server configuration in VS Code and keep the extracted folder in place. The button uses the bundled runtime; Node, npm and a VSIX are not required. This installs MCP tools. Install the shared skill separately if your agent supports it. See the official MCP installation-link documentation.

Install the shared skill through skills.sh

For agents supported by the skills CLI (Node 22.20+):

npx skills add A-Box-of-Tools/abox-tools-plugin --skill abox-local-files

Then ask the agent to set up the abox.tools local adapter. The skill includes a Python 3.9+ setup script that downloads the matching packaged release, verifies its checksum and installs the bundled runtime. It does not require cloning or building this project. See standalone setup. The skills.sh listing is public. Public v0.11.0 Gemini installation and skill copying were verified; a Gemini gallery listing is unconfirmed.

Ask the agent

• “Compress this photo below 500 KB and save a new copy.”

• “Resize this photo to fit within 1200 by 800 pixels.”

• “Convert this PNG to WebP.”

• “Merge these PDFs in this order.”

• “Split this PDF into one file per page.”

• “Extract pages 1–3 and 4–6 as two PDFs.”

• “Make one PDF containing pages 3, 1–2, and another copy of page 3.”

• “Rotate pages 2–4 clockwise by 90 degrees.”

• “Put these photos into one PDF, in this order.”

• “Check whether this image contains EXIF or XMP metadata.”

• “Remove location and other image metadata, keeping its orientation and colors.”

• “Calculate SHA-256 checksums for these files.”

• “Compress this scanned PDF for screen reading.”

• “Add a diagonal DRAFT watermark to every PDF page.”

• “Make an ICO with 16, 32, 48 and 256 pixel versions of this logo.”

• “Export this SVG as a transparent PNG, 1200 pixels wide.”

• “Arrange these photos into a two-column grid with transparent gaps.”

• “Extract the tables on pages 2–4 into a CSV I can review.”

• “Format this JSON without changing large numbers or key order.”

• “Decode this Base64 file into a new binary file.”

• “Convert this YAML configuration into JSON.”

• “Format this XML while preserving mixed text.”

• “Compare these two files and save a unified patch.”

• “Generate a QR code from the exact text in this local payload file.”

• “Make a Code 128 barcode from this local payload file.”

• “Embed this image as a CSS custom property without changing its bytes.”

• “Inspect the frame timings of this GIF without decoding its pictures.”

• “Extract GIF frames 1, 5 and 10 as composited PNGs.”

• “Trace this black-and-white logo into a single-color SVG.”

• “Average these aligned, opaque images into a new PNG.”

• “Black out the exact regions in my local JSON file and save a new PNG.”

• “Generate two six-word passphrases into a new local file without showing them.”

All input and output paths must be absolute. Existing outputs are never overwritten. Image compression accepts an exact byte target; a result reports whether the target was met.

The six new tools' settings and JSON configuration examples are documented in the shared skill reference.

Operations added in v0.9.0

These routes extend the earlier v0.8.0 operation set. Their full settings, local JSON shapes and review requirements are in the shared skill's media and document reference. Use capabilities to check what your installed release exposes.

In v0.9.0, audio inputs are simple mono/stereo PCM16/Float32 WAV. Version 0.10.0 expands the admitted codecs as described below. MP4 creation is limited to 60 seconds, 1800 frames and a 2,073,600-pixel output, and needs a supported installed browser encoder. Receipt OCR includes its worker, WebAssembly and English data; no processing asset is downloaded. These are operation-specific bounds, not a universal process-memory guarantee. EXIF edits use a separate local edits_path with bounded set/remove instructions; see metadata editing. Editing retains the thumbnail and other metadata and does not anonymize an image. See additional text formatting for exact HTML/CSS/YAML options, XML/JSON mapping losses and the passive Markdown dialect. Formatting existing HTML keeps its active content; it is distinct from the Markdown renderer's escaped, restricted output.

Operations added in v0.10.0

These additions are included in the public v0.10.0 release. Always check the installed adapter's capabilities before choosing a new operation.

Audio editing/trimming also accepts bounded MP3 Layer III, AAC-LC ADTS and an ordinary unedited MP4/M4A/MOV AAC-LC track. Compressed WAV output needs a native packet decoder; the adapter never falls back to unbounded whole-file decoding. Check actual decoded duration because priming/padding can reduce sample counts. See audio codecs and extraction.

DICOM now also admits a strict single-frame JPEG Lossless Process 14 subset, with complete entropy admission and independent stored-sample checks. It remains a technical preview and can retain burned-in identifiers. See DICOM details.

Use exact user-provided geometry for PDF redaction; the adapter does not find sensitive regions or erase the original. Searchable text, signatures and interactive PDF features are lost. For image conversion, read HEIC/AVIF limits, including 8-bit SDR output and metadata omission. No route downloads a codec while processing a file.

The seven new video processing routes admit stable progressive 8-bit 4:2:0 H.264 before native decoding; this is not arbitrary codec support. Generated MP4 frames are decoded and checked for shape, timing and count, while copied AAC packets are compared exactly. Lossy encoding, color conversion and final-frame timing can change. Matrix-only video rotation remains available separately without decoding or re-encoding.

Operations added in v0.11.0

These controls open a separate visible browser and keep private content and pairing secrets outside agent context. The person starts/connects and approves capture/peers there. They require a graphical desktop and directly reachable devices; host-candidate connections can require the same network. File operations remain offline. These controls are included in the verified public v0.11.0 release; historical v0.10.0 receipts remain above.

At most four windows run together. ttl_minutes defaults to 30 and ranges from 1 to 120. Share attachments are at most 16 files, 128 MiB each and 256 MiB combined; text is at most 50,000 characters and 200,000 UTF-8 bytes. Peers' names never select filesystem paths: explicit receive saves use generated .bin or .txt names with exclusive writes. The synthetic loopback-looking address has no actual HTTP listener or file server. See live settings and CLI lifetime, Privacy and Security.

For PDF protection, pass open_password_file/owner_password_file; for unlocking, pass password_file. Their exact UTF-8 text, including BOM and final newline, is used without trimming. Never put passwords or decoded code contents into a tool request or conversation. Document scans and ID photos require dimensions and source pixel coordinates from a local JSON file, supplied deliberately by the user. The adapter does not infer sensitive regions or face landmarks.

Use the CLI

The CLI provides the same operations without MCP. These commands use Node in a source checkout. In a release bundle, replace node with ./runtime/node, or .\runtime\node.exe on Windows:

node bin/abox.mjs capabilities
node bin/abox.mjs doctor
node bin/abox.mjs compress-image --request-file /absolute/path/compress.json
node bin/abox.mjs resize-image --request-file /absolute/path/resize.json
node bin/abox.mjs convert-image --request-file /absolute/path/convert.json
node bin/abox.mjs split-pdf --request-file /absolute/path/split.json
node bin/abox.mjs merge-pdf --request-file /absolute/path/merge.json
node bin/abox.mjs select-pdf-pages --request-file /absolute/path/select.json
node bin/abox.mjs rotate-pdf --request-file /absolute/path/rotate.json
node bin/abox.mjs images-to-pdf --request-file /absolute/path/photos.json
node bin/abox.mjs inspect-exif --request-file /absolute/path/inspect.json
node bin/abox.mjs remove-exif --request-file /absolute/path/metadata.json
node bin/abox.mjs hash-file --request-file /absolute/path/hash.json
node bin/abox.mjs compress-pdf --request-file /absolute/path/compress-pdf.json
node bin/abox.mjs watermark-pdf --request-file /absolute/path/watermark.json
node bin/abox.mjs image-to-ico --request-file /absolute/path/icon.json
node bin/abox.mjs svg-to-image --request-file /absolute/path/svg.json
node bin/abox.mjs image-layout --request-file /absolute/path/layout.json
node bin/abox.mjs pdf-to-csv --request-file /absolute/path/tables.json
node bin/abox.mjs format-json --request-file /absolute/path/format.json
node bin/abox.mjs base64 --request-file /absolute/path/base64.json
node bin/abox.mjs image-to-data-uri --request-file /absolute/path/embedding.json
node bin/abox.mjs convert-yaml-json --request-file /absolute/path/conversion.json
node bin/abox.mjs format-xml --request-file /absolute/path/xml.json
node bin/abox.mjs generate-qr --request-file /absolute/path/qr.json
node bin/abox.mjs generate-barcode --request-file /absolute/path/barcode.json
node bin/abox.mjs diff-text --request-file /absolute/path/diff.json
node bin/abox.mjs inspect-gif --request-file /absolute/path/gif-info.json
node bin/abox.mjs split-gif --request-file /absolute/path/gif-frames.json
node bin/abox.mjs trace-image --request-file /absolute/path/trace.json
node bin/abox.mjs stack-images --request-file /absolute/path/stack.json
node bin/abox.mjs redact-image --request-file /absolute/path/redaction.json
node bin/abox.mjs generate-passwords --request-file /absolute/path/passwords.json

A compression request file contains:

{
  "input_path": "/absolute/path/photo.jpg",
  "output_path": "/absolute/path/photo-small.jpg",
  "target_bytes": 500000,
  "format": "keep",
  "allow_resize": true
}

A page-range request file contains:

{
  "input_path": "/absolute/path/document.pdf",
  "output_dir": "/absolute/path/document-parts",
  "mode": "ranges",
  "ranges": ["1-3", "4-6"],
  "keep_bookmarks": true
}

An image metadata request file contains:

{
  "input_path": "/absolute/path/photo.jpg",
  "output_path": "/absolute/path/photo-clean.jpg",
  "keep_orientation": true,
  "keep_icc": true
}

--request '<JSON>' is also supported, subject to your shell's quoting rules. Request files avoid those differences. The CLI prints JSON summaries; it never prints file contents.

Limits and behavior

• Image compression, resizing and conversion: JPEG, PNG, and static WebP. Animated images are refused. Resizing fits the full image (contain) or crops centrally (cover); upscaling is disabled by default. Re-encoding can change colors and removes source metadata. A no-op resize preserves original bytes.

• Metadata removal: JPEG, PNG, and WebP containers, including their animation.

• PDF page operations: unencrypted input, up to 5,000 pages and 1,000 output files. Merging accepts 2–32 PDFs, at most 128 MiB combined. Page selection preserves the order and repetition of comma-separated terms. Ranges expand ascending: 3,1-2,3 produces 3,1,2,3; use explicit terms for reverse order. Merging PDFs with colliding interactive form field names is currently unsupported by the source engine and fails before creating an output. Tool responses include at most 20 output entries. Larger splits also write a local manifest.json containing every output path and its source pages.

• Images to PDF: still JPEG, PNG or WebP, one image per page, up to 128 MiB combined and 80 megapixels per image. Defaults to A4, automatic orientation, contain, no margins. keep can embed original JPEG streams with their metadata; remove metadata first when a clean document is required. lossless preserves decoded image pixels; jpeg re-encodes using the requested quality.

• Checksums: read-only, streaming, supports empty files and up to 8 GiB.

• PDF compression: presets recompress supported images and rewrite structure. smallest uses 96 DPI / 0.55 quality, screen 130 / 0.68, print 220 / 0.82, and gentle keeps resolution with 0.90 quality. dpi and quality can override those settings. It does not search for a byte target; check smaller and actual savings. The output can be larger. Changed images can lose detail. CMYK and unsupported image encodings are retained where safely readable. Unusual filter-wrapped JPEG encodings and nonidentity EXIF orientation in JPEG streams are refused before processing to prevent unintended image rotation. Standard PDF metadata is removed by default; embedded images, attachments and visible content can still contain personal information.

• PDF watermarks: one rasterized line, up to 200 characters, centered or tiled, on all pages or only the first. Local fonts affect appearance. Up to 100,000 stamps. Watermarks can be removed by editors and do not redact or protect data.

• ICO: still JPEG, PNG or WebP to one file with up to ten sizes from 16 to 256 pixels. auto stores sizes up to 64 as DIB and larger sizes as PNG. Padding preserves aspect ratio and transparency; crop fills the square; stretch can distort. The directory, frame dimensions and rendered pixels are verified.

• SVG: uncompressed, self-contained static SVG to PNG, JPEG or WebP; up to 16,384 pixels per side and 80 megapixels. Active content, foreign objects, external references, CSS imports/escapes and animations are refused. JPEG fills transparency with white by default. Fonts and colors can vary by browser.

• Image layouts: 2–32 still JPEG, PNG or WebP images, at most 128 MiB combined and 80 megapixels per input. Output is at most 8192 pixels per side and 32 megapixels. Grids default to two square columns; horizontal/vertical strips preserve original aspect ratios. Contain retains the full image; cover crops. New pixels are rendered in 8-bit SDR and source metadata is not copied.

• PDF to CSV: selectable text only, no OCR. Tables, headings, dates and amounts are inferred and can be incomplete or wrong; review the CSV against the PDF. UTF-8 BOM, comma delimiter, CRLF lines. Multiple tables have a blank separator row. Year-bearing dates and inferred amounts are normalized; auto date order falls back to day-first when ambiguous. Formula-like text is escaped with an apostrophe by default. Up to 1000 tables, 64 columns per table, 100000 data rows, one million cells/glyphs and 16 MiB of output. Structural verification does not establish extraction accuracy.

• PDF table extraction currently accepts unfiltered and Flate streams without predictors. It conservatively refuses legacy LZW/RunLength, ASCII-armoured streams and predictors, even in unused images or xrefs. Blocked PDF names in printed text can also cause refusal. Stream expansion is limited to 16 MiB per Flate stream and 64 MiB cumulatively before buffering decoded data.

• JSON: UTF-8 input up to 8 MiB, nesting up to 256 levels, output up to 32 MiB. Preserves raw number and string tokens, duplicate keys and object key order. Sorting is opt-in. Both pretty and minified output end with a newline. The parser accepts at most 250000 syntax nodes, including object keys.

• Base64: arbitrary binary files, including empty files; input up to 8 MiB, output up to 32 MiB. Standard encoding is padded; URL-safe encoding is unpadded. Relaxed decoding accepts whitespace, missing padding and either alphabet. Canonical validation requires the exact requested spelling and zero unused bits. Encoded/decoded bytes are written locally and never returned to the agent.

• Image data URIs: input up to 8 MiB and output up to 32 MiB. Signature-based MIME recognition does not prove image validity or rendering. Original bytes, metadata and SVG markup are preserved; this is not a sanitizer. The five output shapes use .txt for uri, .css for css-rule/css-var, .html for html, and .md for markdown. Auto encoding uses readable percent encoding only for exact UTF-8 SVG in a standalone URI. SVG embedding envelopes use Base64 to avoid HTML, CSS or Markdown changing the URI's spelling.

• YAML/JSON conversion: UTF-8 input up to 2 MiB, output up to 16 MiB, 128 nesting levels and 100000 syntax nodes. Supports the website's YAML subset, not every YAML feature. Anchors, aliases, custom tags and multiple documents are refused. Ordered values, duplicate keys and large numeric tokens are retained; comments, formatting and string escape spelling can be lost. Optional natural key sorting applies only when converting YAML to JSON. Output must round-trip before it is written. The source printer cannot handle root multiline strings or some four-space lists of mappings; those requests are refused. Use the default two-space indentation for lists of mappings.

• XML formatting: the same 2 MiB input, 16 MiB output, 128 nesting levels and 100000-node limits. UTF-8 only; DTDs and custom entities are refused, without fetching external resources. Predefined entities and valid numeric references are supported. Mixed content and xml:space follow the source engine's preservation rules; formatting may change insignificant whitespace and quotes.

• QR codes and barcodes: read payloads directly from UTF-8 files, including any final newline. Payloads and completed barcode digits are not returned to the agent. QR input is at most 4 KiB; barcode input at most 256 bytes. Actual encoder and verification capacity can be smaller. Retail check digits are computed or validated; invalid supplied digits fail. Code 39 can uppercase text and optionally add its checksum. SVG is the default and needs no browser; PNG uses the installed browser. Output is at most 8192 per side, 32 megapixels and 16 MiB. Barcode verification permits at most 400 alternating runs; plain ITF is limited to 4–40 digits. Check warnings and test the saved code with the intended scanner before printing or relying on it.

• Text comparison: two UTF-8 files, including empty files, at most 4 MiB and 20000 logical lines combined; NUL-containing binary inputs are refused. Writes a new .diff or .patch up to 8 MiB. The patch always describes the original strings, including BOMs, line endings and final newline differences; ignore options affect comparison_stats only. Primary counts describe line content, with terminator changes reported separately. The adapter verifies that patch replay reconstructs the right input from the left. Header labels are fixed a and b. Large differences can produce coarse rewrite hunks; minimality is not guaranteed. Neither source is modified or patched.

• Each other input: at most 128 MiB.

• GIF inspection: 32 MiB, 1000 source frames, 8192 per side and 32 megapixels per screen or patch. Reads headers without verifying LZW or pixels. Frame results are paginated; comments, application text and palette colors are excluded.

• GIF extraction: at most 200 selected PNGs, 128 Mi decoded patch pixels and 300 million walked screen pixels. Each PNG is at most 32 MiB and the total PNG output at most 256 MiB; the upstream working-buffer admission also applies. Strict decoding refuses partial frames. Composited mode applies preceding disposal; stored mode returns patches. Output pixels are checked after PNG encoding, and a local timing manifest lists every frame. Needs a local browser.

• Image tracing: still JPEG/PNG/WebP, at most 4096 per side and 4 megapixels. Single-color, lossy SVG shape approximation for logos or line art; not photo reconstruction. At most 100000 boundary edges, 1000 contours and 400000 SVG bytes. Passive path structure and rendered path pixels are checked.

• Stacking: 2–8 opaque, equal-size still JPEG/PNG/WebP frames, up to 128 MiB combined, 4096 per side and 4 megapixels. No alignment, resizing, focus stacking or RAW development. Uses encoded 8-bit SDR RGB; PNG pixels are lossless after stacking, JPEG is lossy. Both stacking and tracing require the local browser.

• Redaction: explicit 0-based integer rectangles in the decoded display image. Expected dimensions must match exactly; rectangles cannot be clipped. Up to 100 regions, 16384 per side, 32 megapixels and 128 million rectangle pixels including overlaps. Solid black fill only, flattened PNG, with all decoded pixels checked. Covered pixels are exactly opaque black; outside alpha is retained and visible colors are within one 8-bit step of the decoded source. The new raster is 8-bit SDR and excludes source metadata. Does not locate sensitive content; review coverage before sharing. Requires the local browser.

• Passwords: cryptographic generation into a new plaintext .txt, one secret per line, at most 100 results and 64 KiB. Defaults to 24 characters; passphrase mode defaults to six EFF words. Settings come from local JSON up to 16 KiB. Secrets, secret hashes and samples never appear in tool results. No browser or vault.

• Compression: up to 80 megapixels and two minutes per image. auto may adjust the output extension to match the selected format. Other modes require the output extension to match the actual bytes, including an unchanged original.

• Compression may resize an image unless allow_resize is false. A file that already meets the target can pass through unchanged, retaining its metadata. Use metadata removal when that is part of the task.

• Metadata removal can retain animation in PNG and WebP because it edits their containers; compression refuses animation rather than flattening it.

• PDF splitting creates new PDFs. Document tags and metadata may be removed; digital signatures do not survive as valid signatures. Output page counts are verified. Bookmark preservation is enabled by default where supported.

• Metadata removal retains orientation and ICC profiles by default to preserve appearance. Set keep_orientation and keep_icc to false to remove those supported fields too; display orientation or colors may change. It does not establish that an image is anonymous.

Downloading a release or installing a browser uses the network. Processing files through this adapter does not send their bytes over the network. The agent and its host still receive paths, settings, summaries, and errors. The adapter needs no account or API key.

See the plugin privacy notice for the data handled by the adapter and the separate responsibilities of the agent host.

The compressor detects common browser installations. Set ABOX_BROWSER_PATH to an existing browser executable if needed. The browser's sandbox must work on the host OS; Linux detection prefers installed stable Chrome over generic Chromium builds. The plugin does not download a browser during processing.

This plugin and the website's development MCP

This plugin runs file operations locally over MCP stdio. The separate website catalogue at mcp-dev.abox.tools/mcp exposes find_browser_tools: it accepts a public search query, locale, and limit, then returns tool links and an interactive catalogue. It accepts no files or filenames and does not process files. Its source is the website's development catalogue worker.

The mcp-dev endpoint is for development. This repository's release workflow does not deploy the website or publish a production catalogue endpoint. For directory submissions and MCP registration, see Publishing. For Gemini CLI, Vercel skills.sh, Cursor and other distribution options, see Platform support and publishing. For additional directories, local distribution and their account or licensing requirements, see Other sharing platforms.

Developer setup and other MCP hosts

Use the release bundle for installation without development tools. To develop from a source checkout, install Node.js 22.15 or later, then run:

npm ci --ignore-scripts
npm test
npm run test:browser
npm run doctor

Source setup downloads npm dependencies. The source checkout's mcp.json uses Node from the host; release packaging replaces it with the bundled executable. A developer can add this repository marketplace with codex plugin marketplace add A-Box-of-Tools/abox-tools-plugin, then install it in the desktop app and run Setup in the installed copy. See the official OpenAI packaging guide.

For another host supporting local MCP stdio, point its configuration at the extracted bundle's executable and server script using absolute paths:

{
  "mcpServers": {
    "abox": {
      "command": "/absolute/path/to/bundle/runtime/node",
      "args": ["/absolute/path/to/bundle/bin/abox-mcp.mjs"]
    }
  }
}

On Windows, use runtime/node.exe. Escape backslashes in JSON or use forward slashes. A source checkout can use node instead of the bundled executable. Follow the host's own configuration format; do not replace unrelated settings.

Maintain the reused code

To refresh the pinned website sources from a local website checkout:

npm run sync-upstream -- --source /absolute/path/to/etoolbox

Review the vendor diff and provenance, retain license notices, and run the checks above. A v* tag starts .github/workflows/release.yml: it tests, builds all four platform bundles, builds a universal MCPB, and checks extracted MCP execution on all four platforms before publishing the GitHub Release. A successful tagged release then publishes its verified metadata to the Official MCP Registry using GitHub's short-lived identity token. Claude and OpenAI directory reviews remain separate submissions; see Publishing.