Pixie is one local web app. Point it at a tools/ folder and it scans, validates, and serves every tool in there with a real UI: forms, charts, maps, tables, audio, video, the lot. Each tool runs in its own uv virtualenv as a subprocess that spins up on click and shuts down when idle.
You don't add tools through the dashboard. You add them by talking to Claude Code, and you bring whatever you've already got:
- a description in plain English (
add-tool-from-description) - a GitHub repo (
add-tool-from-repo) - a Python script on disk (
wrap-local-script) - a Jupyter notebook (
add-tool-from-notebook) - a Streamlit or Gradio app (
convert-streamlit-app,convert-gradio-app) - a CLI binary like
ffmpeg,yt-dlp,pandoc(add-tool-from-cli-command) - an OpenAPI / Swagger spec (
add-tool-from-openapi-spec) - an Excel
.xlsxmodel, formulas and all (add-tool-from-excel-model) - an academic paper PDF or arXiv link (
add-tool-from-paper, which also generates reference fixtures from the paper's reported numbers) - a detailed model spec when there's no paper or code (
implement-model-from-spec)
Each skill writes the schema, the FastAPI handler, the dependencies, runs the validator end-to-end, and only claims success if every check passes. Refresh the page and the tool shows up with a green dot.
Everything stays on 127.0.0.1. No accounts. No telemetry. No cloud sync. Your .env files stay in your tool folders.
git clone https://github.com/AlexKapadia/Pixie
cd Pixie
uv sync
uv run pixieOpen http://127.0.0.1:7860. Done.
Requires Python 3.12 and uv. Works on macOS, Linux, and Windows.
In Claude Code, inside the Pixie repo, say whichever fits what you've got:
Add a Pixie tool that fetches a stock ticker and runs a Monte Carlo on it.
Add a Pixie tool from https://github.com/some-org/cool-model
Wrap my Excel model at ./quant/black_litterman.xlsx as a Pixie tool.
Implement this paper as a Pixie tool: https://arxiv.org/abs/2310.06770
Add a Pixie tool from this notebook: ./experiments/forecasting.ipynb
Claude picks an id, writes tool.json, writes main.py, runs uv sync, spawns the subprocess, POSTs sample input, checks the response conforms to the declared output schema, then reports back. If anything fails, the tool does not appear in the sidebar.
Full walkthrough: start/first-tool.
The runtime, the validator, the schema-driven UI, and the full Claude Code skill library — nothing else. tools/ is empty on a fresh clone. You bring your own scripts, notebooks, repos, papers, or Excel models and Claude wires them in.
Every tool lives in its own folder under tools/ with a tool.json, a pyproject.toml, a main.py, sample fixtures, and reference outputs the validator diffs against on every change. Local .venv, data/, models/, outputs/, and .env stay on your machine and never get committed.
tool.json declares inputs, outputs, layout, and secrets. Pixie renders the entire UI from the schema, no per-tool frontend code.
{
"id": "currency-converter",
"name": "Currency Converter",
"inputs": [
{ "id": "from", "type": "select", "options": ["USD", "EUR", "GBP"] },
{ "id": "to", "type": "select", "options": ["USD", "EUR", "GBP"] },
{ "id": "amount", "type": "number", "min": 0 }
],
"outputs": [
{ "id": "converted", "type": "number", "format": "currency" }
]
}You get, free, 28 input widgets and 30-odd output renderers: text, markdown, sliders, file drops, maps (points, heatmap, choropleth, polygons, route), charts (line, bar, scatter, area, pie, histogram, boxplot, heatmap, candlestick, radar, sankey, treemap, network), tables, audio, video, LaTeX, code, diffs, gantt, timelines, streaming text.
Full reference: Input types and Output types.
Pixie ships about fifty Claude Code skills covering the whole lifecycle:
- Build new tools.
add-tool-from-description,add-tool-from-repo,add-tool-from-paper,add-tool-from-notebook,add-tool-from-openapi-spec,add-tool-from-excel-model,add-tool-from-cli-command,convert-streamlit-app,convert-gradio-app,wrap-local-script,implement-model-from-spec. - Maintain existing tools.
update-tool,rename-tool,fork-tool,tag-tool,migrate-tool-format,organise-tool,lint-tool,set-secret,cite-source. - Diagnose.
debug-tool,pixie-doctor,pixie-status,revalidate-all,validate-against-reference,view-logs,view-runs,inspect-tool,list-tools. - Data and outputs.
fetch-dataset-from-url,fetch-dataset-from-kaggle,fetch-dataset-from-huggingface,import-dataset-from-local,list-outputs,find-output,export-as-format,export-run,export-run-as-report,bulk-export,copy-to-clipboard,copy-output-to,send-to-folder,register-export-target,star-run,label-run. - House-keeping.
archive-tool,unarchive-tool,remove-tool,audit-disk-usage,clear-old-outputs,share-tool,import-tool,workspace-create,workspace-add-tool,workspace-remove-tool,open-artefacts-folder.
Each one has a narrow trigger, runs the validator if it touches a tool, and refuses politely when given the wrong shape of input.
Every tool passes a deterministic check sequence before it appears in the sidebar:
- Folder structure is sane.
tool.jsonparses against the schema.pyproject.tomlis valid..venvbuilds cleanly.- Subprocess starts on the assigned port.
/healthzreturns 200./schemamatchestool.json./runaccepts sample input and returns inside the timeout.- Output shape matches the declared outputs.
- Subprocess shuts down cleanly when idle.
- Reference fixtures diff inside tolerance.
Failures don't ship. There is no "mostly works" tier.
Pixie is not a SaaS, not a marketplace, not a deployment platform. It does not authenticate users. It does not push tools to a server. It will not, ever, ingest a tool from inside its own UI: that path goes through Claude Code on your machine, where you can see what is being written before it is written.
It is also not a generic app platform. It is for small tools and models with a clean input/output contract. If you need session state, multi-user auth, or long-lived background jobs, build them somewhere else.
The Pixie test suite (unit, integration, format, perf, visual regression, accessibility, e2e Playwright matrix) lives in a separate repo: AlexKapadia/Pixie-tests. Keeps this repo focused on the runtime and the tool library.
Active, single-author project, MIT licensed. Issues and PRs welcome at github.com/AlexKapadia/Pixie/issues. Read Contributing before opening a PR, but the short version is: if it works and the validator passes, it's probably in.
MIT.