Alpha

WorkBench is an AI Workbench built to bring the most powerful models from across the web into a seamless user experience.

WorkBench

Run JavaScript

Run general-purpose JavaScript in Toolbox. Recommended for document generation, image processing, archives, and file creation. Choose the runtime whose listed libraries best support the task. Prefer available libraries over reimplementing their functionality. Runtime recommendations are not restrictions. Always try to reuse an existing session_id for related work. If the response says sessions are unavailable, each call starts fresh; include all required inputs in every call. Save reusable values in SESSION_STATE; ordinary const and let variables are local to each call. Responses list saved variables and loaded libraries. Select libraries with the libraries parameter: papaparse (Papa) for CSV/TSV, xlsx (XLSX) for spreadsheets, lodash (_) for data transformations, jszip (JSZip) for ZIP archives, pdf-lib (PDFLib) for creating/editing PDFs, docx (docx) for Word documents, pptxgenjs (PptxGenJS) for creating presentations, jimp (Jimp) for raster images, and ffmpeg for audio/video processing. Selected libraries remain loaded in the session. Omit libraries or pass [] when no additional libraries are needed. Use the selected globals directly; no imports or package installation are needed. This is a browser sandbox: use in-memory bytes and export through output.file, not filesystem or download helpers. Keep Papa worker and download options disabled and check parse errors. pdf-lib cannot extract text or render PDFs; standard PDF fonts have limited character support. pptxgenjs cannot import existing presentations. Jimp uses the v1 API: fromBuffer(bytes), resize({ w, h }), crop({ x, y, w, h }), and new Jimp({ width, height, color }). Use getBuffer for image export; avoid filesystem paths, URL loading, and font downloads. Use browser Canvas for text drawing and SVG/WebP decoding. Selecting ffmpeg provides an already-loaded ffmpeg instance; do not call load() or terminate(). Write inputs with await ffmpeg.writeFile("input.mp4", new Uint8Array(await files[0].arrayBuffer())). Run await ffmpeg.exec(["-i", "input.mp4", "output.mp3"], 60000); check the returned exit code is 0, then export await ffmpeg.readFile("output.mp3") through output.file. For ffmpeg.ffprobe, write JSON metadata to a file with -o and validate that file; this build may return -1 even on success. FFmpeg and its virtual files persist in the session; delete unneeded files with await ffmpeg.deleteFile(name). Prefer short clips and small files; transcoding is CPU- and memory-intensive. All FFmpeg assets are bundled, so local-file processing works without network access. Selected attachments are available in the files array as File objects, in the requested order, each with an additional id property. Read UTF-8 text with await files[0].text(), binary bytes with await files[0].arrayBuffer(), or raster images with await createImageBitmap(files[0]). For SVG, use new Image() with an image/svg+xml Blob URL created inside the sandbox, await image.decode(), and draw the image onto a canvas; do not pass an SVG Blob to createImageBitmap. Include the SVG namespace and explicit width/height in SVG markup. Revoke the Blob URL after drawing. Export PNG with await canvas.convertToBlob({ type: "image/png" }) on an OffscreenCanvas. When network access is disabled, only blob: and data: image sources are allowed. Return a value to produce text output; objects and arrays are serialized as JSON. Create chat attachments with output.file({ filename, data, mimeType }); data accepts a string, Blob, File, ArrayBuffer, or typed array. Use a filename without paths. mimeType is optional and defaults to text/plain for strings, the Blob type when available, or application/octet-stream. Call output.file multiple times to create multiple files; you can also return text in the same run. Network access is disabled by default and controlled by the Toolbox setting.

toolbox_run_js

Run Python

Run general-purpose Python in Toolbox. Recommended for data processing and analytics, including dataset cleaning, aggregation, statistics, and numerical calculations. Choose the runtime whose listed libraries best support the task. Prefer available libraries over reimplementing their functionality. Runtime recommendations are not restrictions. New sessions provide the Python standard library. Import its modules normally. Load third-party dependencies through the required packages parameter before code runs. Available packages: numpy for numerical arrays, pandas for tabular data analysis, matplotlib and seaborn for static charts, plotly and bokeh for interactive HTML charts, pillow for images, openpyxl for Excel workbooks, pypdf2 for reading and editing PDFs, python-docx for Word documents, requests for HTTP requests, beautifulsoup4 for HTML parsing (from bs4 import BeautifulSoup), jinja2 for text/HTML templates, and jsonschema for JSON validation. Dependencies load automatically, including NumPy, pandas, Matplotlib, and Pillow where required. Write packages before code. For import pandas as pd, include "pandas" unless the session response already lists it as loaded. Use packages: [] for standard-library-only code or when all required packages are already loaded in the session. Then import the selected packages in code; loading a package does not create an import alias. If an available package raises ModuleNotFoundError, retry in the same session with that package included in packages. Other third-party packages and pip installation are unavailable. For Matplotlib or seaborn charts, call matplotlib.use("Agg") before importing matplotlib.pyplot as plt or seaborn, and save figures to io.BytesIO. Pass the buffer's getvalue() to output.file with a filename and MIME type; close figures with plt.close(fig). For Plotly, export fig.to_html(include_plotlyjs=True, full_html=True). For Bokeh, use bokeh.embed.file_html with bokeh.resources.INLINE. Export that HTML through output.file with MIME type text/html; avoid show() and CDN resources. Plotly static image export is unavailable. For images, select pillow and import from PIL, for example from PIL import Image. Save images to io.BytesIO with an explicit format, then export its getvalue() through output.file. For pypdf2, import PdfReader and PdfWriter from PyPDF2. For python-docx, import Document from docx. Save Excel, PDF, and Word documents to io.BytesIO and export its getvalue() through output.file. For HTTP, select requests and import requests. Network access must be enabled; browser CORS rules apply. Use a timeout and call response.raise_for_status() before reading the response. For HTML, use BeautifulSoup(markup, "html.parser"). Always try to reuse an existing session_id for related work. If the response says sessions are unavailable, each call starts fresh; include all required inputs in every call. Variables, imports, loaded packages, and virtual files persist in that session. Responses list saved variables and loaded packages. Attachments selected for the current call appear in the files list in the requested order as dictionaries with id, name, type, and path. Their bytes are available in the virtual filesystem at files[0]["path"]. Read text with pathlib.Path(files[0]["path"]).read_text(), bytes with .read_bytes(), or CSV with pd.read_csv(files[0]["path"]). Use print() or a final expression for text output; values use their Python string representation. Create chat attachments with output.file(filename, data, mime_type=None). Data must be a string, bytes, bytearray, or memoryview; filenames must not contain paths. Text defaults to text/plain and bytes to application/octet-stream. For example, output.file("summary.csv", df.to_csv(index=False), "text/csv"). For a file written by a library, read its bytes and pass them to output.file. Call output.file multiple times to create multiple files. Only files passed to output.file become chat attachments. Top-level await is supported, but top-level return is not. Network access is disabled by default and controlled by the Toolbox setting.

toolbox_run_python

Toolbox

Seedream 4.5 is the latest in-house image generation model developed by ByteDance. Compared with Seedream 4.0, it delivers comprehensive improvements, especially in editing consistency, including better preservation of subject details, lighting, and color tone. It also enhances portrait refinement and small-text rendering. The model’s multi-image composition capabilities have been significantly strengthened, and both reasoning performance and visual aesthetics continue to advance, enabling more accurate and artistically expressive image generation.

ByteDance Seed: Seedream 4.5

FLUX.2 [max] is the new top-tier image model from Black Forest Labs, pushing image quality, prompt understanding, and editing consistency to the highest level yet.

Black Forest Labs: FLUX.2 Pro

Recraft V4 Pro is an image generation model from Recraft. It supports text and image inputs with image output at ~2K resolution across multiple aspect ratios, double the resolution of V4. It offers higher fidelity and detail density than V4, suited for production use cases where image quality is a priority.  Supports the following image_config parameters: strength (controls how much the output deviates from the source image), rgb_colors (sets a color palette), and background_rgb_color (sets the background color). See the image generation docs for details: https://openrouter.ai/docs/features/multimodal/image-generation

Recraft V4 Pro

Recraft V4.1 Pro Vector is the vector (SVG) variant of Recraft V4.1 Pro, tuned for high aesthetics. It supports text and image inputs and produces higher-resolution SVG image output across multiple aspect ratios, with typical generation around 20 seconds. Output scales cleanly, making it suitable for icons, logos, and other graphics. V4.1 brings more personality to text and illustrations, smoother gradients, and stronger short-prompt adherence compared to V4 Pro.  Suited for higher-resolution illustration work and production graphics where output should be designed rather than photographed. Supports the following image_config parameters: strength (controls how much the output deviates from the source image), rgb_colors (sets a color palette), and background_rgb_color (sets the background color). See the image generation docs for details: https://openrouter.ai/docs/features/multimodal/image-generation(opens in new tab)  Note: only one input image is supported.

Login to WorkBench