Skip to main content

The paintbrush for Claude. A visual output surface for AI agents — diagrams, tables, images, and interactive controls.

Project description

lux

A visual output surface for AI agents.

License CI PyPI Python

Lux gives AI agents a window they can draw into. It runs an ImGui display server on the local machine, connected by Unix socket IPC. Agents send JSON element trees via MCP tools; the display renders them at 60fps. The protocol is the API surface --- if an agent can describe it as JSON, Lux renders it.

The design follows Smalltalk's Morphic model: every visible element is a composable, nestable object. Windows contain tabs, tabs contain groups, groups contain buttons and plots. The long-term goal is a live environment where the MCP server is the message bus and Lux is the rendering layer, with the agent as the programmer at the keyboard.

Platforms: macOS, Linux

Stage: alpha --- protocol is stable, published on PyPI as punt-lux

A Claude Code plugin displaying a project issue board --- the agent reads JSONL, builds a filterable table with detail panel, and renders it in a single tool call. Filters and row selection run at 60fps with zero MCP round-trips.

Beads issue board with filterable table and detail panel

The same list/detail pattern generalizes to any tabular data. Search, combo filters, pagination, and a detail panel --- all driven by a single show_table() call.

Filterable data explorer with detail panel

Dashboards compose metric cards, charts, and tables. show_dashboard() builds the layout from structured data --- no manual element positioning needed.

Dashboard with metrics, charts, and data table

Quick Start

curl -fsSL https://raw.githubusercontent.com/punt-labs/lux/8aab4ad/install.sh | sh

Restart Claude Code twice. The Lux display window opens automatically when agents send visual output.

Manual install (if you already have uv)
uv tool install punt-lux

Then install the plugin via the marketplace:

claude plugin marketplace add punt-labs/claude-plugins
claude plugin install lux@punt-labs
Verify before running
curl -fsSL https://raw.githubusercontent.com/punt-labs/lux/8aab4ad/install.sh -o install.sh
shasum -a 256 install.sh
cat install.sh
sh install.sh
Run a demo
lux display &
uv run python demos/dashboard.py

Demos are in demos/ --- each connects as a client and drives the display:

Demo What it shows
interactive.py Sliders, checkboxes, combos, text inputs, color pickers
containers.py Windows, tab bars, collapsing headers, groups
dashboard.py Multi-window layout with draw canvases and live controls
data_viz.py Tables, plots, progress bars, spinners, markdown
menu_bar.py Custom menus, event handling, periodic refresh

Features

  • 22 element kinds --- text, buttons, images, sliders, checkboxes, combos, inputs, radios, color pickers, selectables, trees, tables, plots, progress bars, spinners, markdown, draw canvases, groups, tab bars, collapsing headers, windows, separators
  • Layout nesting --- windows contain tab bars contain groups contain any element, arbitrarily deep
  • Incremental updates --- update patches individual elements by ID without replacing the scene
  • Menu bar --- built-in Lux/Theme/Window menus, plus agent-extensible custom menus via set_menu
  • Interaction events --- button clicks, slider changes, menu selections queue as events the agent reads via recv
  • Auto-spawn --- LuxClient starts the display server on first connection if it isn't running
  • Unix socket IPC --- length-prefixed JSON frames, no HTTP overhead, no threads

MCP Tools

Agents interact with Lux through eight MCP tools exposed by lux serve:

Tool What it does
show(scene_id, elements) Replace the display with a new element tree
show_table(scene_id, columns, rows) Display a filterable data table with optional detail panel
show_dashboard(scene_id, ...) Display a dashboard with metric cards, charts, and a table
update(scene_id, patches) Patch elements by ID (set fields or remove)
set_menu(menus) Add custom menus to the menu bar
set_theme(theme) Switch display theme (dark, light, classic, cherry)
clear() Remove all content from the display
ping() Round-trip latency check
recv(timeout) Read the next interaction event (clicks, changes)

What It Looks Like

Show text and a button

{"tool": "show", "input": {
  "scene_id": "hello",
  "elements": [
    {"kind": "text", "id": "t1", "content": "Hello from the agent"},
    {"kind": "button", "id": "b1", "label": "Click me"}
  ]
}}

Returns "ack:hello". When the user clicks the button:

{"tool": "recv", "input": {"timeout": 5.0}}

Returns "interaction:element=b1,action=click,value=True".

Multi-window dashboard

{"tool": "show", "input": {
  "scene_id": "dash",
  "elements": [
    {"kind": "window", "id": "w1", "title": "Controls", "x": 10, "y": 10,
     "children": [
       {"kind": "slider", "id": "vol", "label": "Volume", "value": 50}
     ]},
    {"kind": "window", "id": "w2", "title": "Chart", "x": 320, "y": 10,
     "children": [
       {"kind": "plot", "id": "p1", "title": "Trend",
        "series": [{"label": "y", "type": "line",
          "x": [1,2,3,4], "y": [10,20,15,25]}]}
     ]}
  ]
}}

Update a single element

{"tool": "update", "input": {
  "scene_id": "dash",
  "patches": [
    {"id": "vol", "set": {"value": 75}}
  ]
}}

Element Kinds

Category Kinds
Display text, button, image, separator
Interactive slider, checkbox, combo, input_text, radio, color_picker
Lists selectable, tree
Data table, plot, progress, spinner, markdown
Canvas draw (line, rect, circle, triangle, polyline, text, bezier)
Layout group, tab_bar, collapsing_header, window

All elements with an id support an optional tooltip field (string shown on hover).

CLI Commands

Command What it does
lux display Start the display server (ImGui window)
lux serve Start the MCP server (stdio transport)
lux status Check if the display server is running
lux version Print version

Architecture

Agent (Claude Code)
  │ MCP (stdio)
  ▼
lux serve (FastMCP)
  │ Unix socket (JSON frames)
  ▼
lux display (ImGui + OpenGL)
  │ renders at 60fps
  ▼
Window on screen

The display server and MCP server are separate processes. The MCP server is a thin adapter that translates MCP tool calls into protocol messages sent over the Unix socket. The display server runs the ImGui render loop, polls the socket each frame via select() with zero timeout, and renders whatever scene the agent last sent.

Client code can also use LuxClient directly as a Python library, bypassing MCP. The demos do this.

Documentation

Architecture (PDF) | Design Log | Changelog | Contributing

Development

uv sync --extra dev            # Install dependencies
uv run ruff check .            # Lint
uv run ruff format --check .   # Check formatting
uv run mypy src/ tests/        # Type check (mypy)
uv run pyright                 # Type check (pyright)
uv run pytest                  # Test

Acknowledgements

Lux is a thin orchestration layer. The rendering is done by Dear ImGui, Omar Cornut's immediate-mode GUI library. ImGui handles all the hard problems --- text layout, widget state, input handling, GPU rendering --- and does so in a single-pass retained-mode-free architecture that maps naturally to Lux's "send JSON, render this frame" model. The 60fps render loop, the composable widget tree, and the ability to drive a full UI from a socket with no threading are all consequences of ImGui's design.

Python bindings come from imgui-bundle by Pascal Thomet, which packages ImGui, ImPlot, and several other ImGui extensions into a single pip-installable wheel with complete type stubs. imgui-bundle is what makes "install one Python package, get a GPU-accelerated UI" possible.

FastMCP provides the MCP server layer.

License

MIT

Project details


Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

punt_lux-0.7.1.tar.gz (56.6 kB view details)

Uploaded Source

Built Distribution

If you're not sure about the file name format, learn more about wheel file names.

punt_lux-0.7.1-py3-none-any.whl (60.4 kB view details)

Uploaded Python 3

File details

Details for the file punt_lux-0.7.1.tar.gz.

File metadata

  • Download URL: punt_lux-0.7.1.tar.gz
  • Upload date:
  • Size: 56.6 kB
  • Tags: Source
  • Uploaded using Trusted Publishing? Yes
  • Uploaded via: twine/6.1.0 CPython/3.13.7

File hashes

Hashes for punt_lux-0.7.1.tar.gz
Algorithm Hash digest
SHA256 2bbd237cfa299793d8cf97afbe42ee98c63888a39a0dd1752c098b561aa2cb82
MD5 6c416e8e218d2fe9bd02eb3101dca863
BLAKE2b-256 fe5a54660ccf1811ad29579472ebe0e59b2f3bc621d061cb96f7f0f0c3666826

See more details on using hashes here.

Provenance

The following attestation bundles were made for punt_lux-0.7.1.tar.gz:

Publisher: release.yml on punt-labs/lux

Attestations: Values shown here reflect the state when the release was signed and may no longer be current.

File details

Details for the file punt_lux-0.7.1-py3-none-any.whl.

File metadata

  • Download URL: punt_lux-0.7.1-py3-none-any.whl
  • Upload date:
  • Size: 60.4 kB
  • Tags: Python 3
  • Uploaded using Trusted Publishing? Yes
  • Uploaded via: twine/6.1.0 CPython/3.13.7

File hashes

Hashes for punt_lux-0.7.1-py3-none-any.whl
Algorithm Hash digest
SHA256 6a205be8ddd120ec63dc4ef55d20a8715e11911a3ff10c3929f9e401ca68ff23
MD5 c5b5942030c978c7d42dcec07b938298
BLAKE2b-256 fb7b0f84e94ba87377b844ccca3accb74039e8103bf9d9c81412d65df8ec0f92

See more details on using hashes here.

Provenance

The following attestation bundles were made for punt_lux-0.7.1-py3-none-any.whl:

Publisher: release.yml on punt-labs/lux

Attestations: Values shown here reflect the state when the release was signed and may no longer be current.

Supported by

AWS Cloud computing and Security Sponsor Datadog Monitoring Depot Continuous Integration Fastly CDN Google Download Analytics Pingdom Monitoring Sentry Error logging StatusPage Status page