Easel

Can an LLM paint with real brush strokes, rather than draw with pixels? Yes — this is the API for it.

Easel is a headless painting engine for AI agents: brushes, paint load, wet blending and canvas texture, driven from Python, from a shell, or over MCP. A brush carries a finite load of paint and runs out along a stroke. Paint lands wet and mixes with what is already there, in a pigment model where blue and yellow make green rather than grey. The canvas has tooth, and a brush low on paint catches only the high points — so dry brush is not a special effect, it is what happens when you run out of paint on rough canvas. Between strokes you look at your own work, which is the whole point: the engine is built for an agent that can see what it just did.

It is not a drawing library, not a rasteriser, and not an image generator — nothing here turns a prompt into a picture. You choose and make every mark. There are no layers, and no undo that costs nothing: you work in passes, and when something is wrong you paint over it.

from easel import Session, blob, cell

s = Session(1024, 768, texture="linen", ground="toned_grey", seed=7)
s.palette["shadow"] = s.palette.mix("ultramarine", "burnt_umber", 0.4)

s.block_in(blob(cell("D5")), brush="bristle", color="shadow", density=0.8)
s.look(values=True)                       # check the value structure
s.stroke([(0.2, 0.6), (0.6, 0.55), (0.9, 0.62)], "bristle", "yellow_ochre")
s.export("painting.png")

Two paintings, made this way

Inside a car wash, from the driver's seat — 206 strokes of a 300 budget, 1152×720 linen, no reference photograph. The nineteen pass scripts beside it reproduce that PNG byte for byte.

Three pears on a kitchen windowsill — 224 strokes, 1024×768 linen, no reference photograph.

Both were painted by a language model working from PAINTER.md alone, one call to this API at a time, with no human hand on the canvas and nothing traced. Every stroke is in the log, both time-lapses were rebuilt from it, and the notes beside each painting say what went wrong as well as what went right. PAINTINGS.md gathers the record, and The worked examples below says what each one is made of.

If you are about to paint from the guide yourself, skip the two pictures above. A worked example names a subject and a named subject leaks: six of six fresh sessions once painted a noun the guide had merely listed. They are here for a reader deciding whether the engine can do this at all, which is a different question from what to paint.

The brushes themselves

Every brush, size and pressure profile, on each canvas texture. Regenerate with python scripts/make_brush_sampler.py — this sheet is the project's primary test artefact, and looking at it catches what the test suite cannot.

A mass does not have to be a rectangle. Each row is one way of building a shape, each column a way of sweeping it; the last column is the box that mass would have been. Regenerate with python scripts/make_shape_sampler.py.

Install

Requires Python 3.12 or newer. Only numpy and Pillow — nothing that is painful to build on Windows.

pip install easel-paint          # the engine, the CLI and the Python API
pip install "easel-paint[mcp]"   # and the MCP server

From a checkout, pip install -e . and pip install -e ".[mcp]" do the same two things.

Either installs an easel command. Pip puts it in the interpreter's scripts directory, which is often not on PATH (it warns when it is not), so python -m easel ... is always available as the same command by another name.

The MCP server is an opt-in extra because nothing else in the engine imports it. See MCP server below.

If you are an LLM agent, read PAINTER.md

PAINTER.md is the guide written for you. It teaches the workflow — tone the ground, paint back to front, check values, refine, edges, highlights last — rather than listing functions, and it opens with the first hour: the whole method on one page, so the eight warm-up exercises come before the long read rather than after it. The facts it would otherwise have to stop and list — units, defaults, what each argument does — are on one page in REFERENCE.md, and the measured numbers behind its rules (graphite survival, wetness decay, the value floor, load windows) are kept apart in CALIBRATION.md, so the guide stays a guide and both can change when the engine does. The engine is designed around one habit:

Look every five to fifteen strokes. A stroke you did not look at was a guess.

llms.txt is the same signpost in the format a model fetching this repository is increasingly told to look for: the summary, what to know before reading further, and where each document is, in about 850 words.

What is in the box

Coordinates are always normalised 0.0–1.0 with the origin top-left. Raw pixels are never exposed, because absolute pixel coordinates are exactly what a language model is worst at.

Command line

Session state lives in a single .easel file, so you can work in increments from a shell without holding a Python process open.

easel new painting.easel --size 1024x768 --texture linen --ground toned_grey --seed 7 --budget 300
easel run painting.easel first_pass.py
easel run painting.easel first_pass.py --rehearse   # against a copy, committing nothing
easel look painting.easel --grid
easel look painting.easel --values
easel look painting.easel --region D4 --fine --reference ref.jpg
easel mark painting.easel top_l 0.335 0.315
easel compare painting.easel ref.jpg
easel prepare painting.easel ref.jpg --level coarse
easel undo painting.easel 3
easel export painting.easel painting.png
easel timelapse painting.easel painting.gif --every 3 --scale 240
easel brushes

Every one of these also works as python -m easel ..., for when the easel executable is not on PATH.

A script run by easel run gets the session pre-bound as s, with the whole public API already in scope — it needs no imports. A prelude.py beside the session file is run first in the same scope, so helpers and mixtures survive between passes; --prelude other.py names a different one and --no-prelude turns it off.

MCP server

The same verbs again, for a client that speaks MCP — and the difference worth having is that the looking tools hand back the picture rather than a path to it. look, preview, rehearse, compare and prepare return their PNG inline, so the loop the guide asks for (look every five to fifteen strokes) costs one call.

pip install "easel-paint[mcp]"
easel-mcp --dir ~/paintings          # or: python -m easel.mcp_server

Or without installing anything, which is the form a client configures:

uvx --from "easel-paint[mcp]" easel-mcp --dir ~/paintings

Both halves of that are named because both are needed: the server is an opt-in extra, and its console script is easel-mcp rather than easel-paint. The server is listed in the official MCP registry as io.github.Gemberkoekje/easel; server.json at the repository root is what is published there.

Fourteen tools: the eleven CLI verbs, plus preview, rehearse and cost — the three questions about a mark that has not been made yet. Marks are made by run, which takes the script as text, and run(rehearse=true) tries a whole pass against a copy and commits nothing. A place is a name, a cell, a span, a rectangle, an outline, or a shape builder like {"blob": "D5", "radius": 0.12}.

run executes Python sent by its client, exactly as easel run does: launch it for a painter you would hand a shell to.

Determinism

Every session takes a seed, and the same script with the same seed produces the same PNG. Each stroke draws its randomness from a generator derived from (seed, stroke index) rather than from one running stream, so a stroke's jitter depends only on which stroke it is — not on how much randomness earlier calls happened to consume.

That is what makes replay exact:

s.replay()          # rebuilds the whole painting from its log; identical export
s.replay(upto=40)   # the state after the first 40 records

It is also how easel undo works across separate shell invocations: session files carry the log, not undo snapshots.

Golden-image tests hold this honest. tests/golden/ stores a hash and a PNG for a fixed script of marks on each texture, plus the whole brush sampler; a change to what a mark looks like fails the suite, and the failure hands you both images to compare. They caught a real one on their first run: the tip-mask cache was keyed on the rounded radius while the mask was built from the exact one, so what a script painted depended on what had run before it in the same process.

Design notes

A few decisions worth knowing about, because they are the ones that make output look painted rather than generated:

• Pigment mixing, not RGB averaging. Colours combine in Kubelka-Munk K/S space using a power mean (p = 0.35). Plain RGB averaging turns every mixture grey-brown; a hard reflectance floor is also needed, or a channel-zero colour swamps the mix and red + blue comes out green. That floor clips the arithmetic only and is taken back off the mixture, so a colour darker than it still lays as written. See src/easel/color.py.
• Spacing is measured along travel. A flat or knife tip is thin in the direction it moves. Spacing it like a round tip leaves a picket fence of discrete bars.
• Wobble is smoothed, not per-dab. Independent per-dab jitter makes neighbouring dabs clump and gap, which reads as banding at the dab frequency. A slow wander along the stroke gives the uneven edge of a real brush without the ripple.
• The tooth gate is roughened with aperiodic grain. Gating a near-periodic weave with a smooth threshold produces a halftone dot screen as paint runs out, which reads as print rather than as dry brush.
• The bristle comb is drawn per stroke, and a bristle has a width of its own. One fixed comb per brush means every wide mark prints the same streaks and a mass laid in passes comes out as corduroy; a fixed count across the tip means the streaks scale with the brush, so a big mass prints stripes wider than anything in the picture and a small mark carries the brush's signature instead of the feature's. So spacing, phase and the missing bristles are redrawn each stroke, and the count follows the brush's size. The same reasoning reaches the round tips through tip_wobble, which is off by default: a disc is the right silhouette for most marks and the wrong one for fifteen small marks in a row, where it prints one shape fifteen times.
• Width follows pressure on the round tips. Pressure that changes only how much paint lands is invisible once an opaque colour saturates, and it means a mark that tapers — a lid, a brow, a lash, a twig — is two strokes at two sizes. The oriented tips keep their chisel, because a flat brush's width is the mass it lays.

Status

Early, and feature-complete against what it was specified to be. The engine, palette, composition helpers, look(), history, CLI, the precision tools (drawing, landmarks, preview, rehearse, compare, prepare), shaped masses and the MCP server all work. The server came last, on the rule that anything changing the API lands before the thing that exposes it, and it has kept up: paint, scumble, cover, circle, union, at_value, the stroke budget, comparison against a written value plan, and rehearsing a whole pass from the shell all arrived in one round after a painter used the guide and wrote down what the engine had cost them.

A second painter did the same thing and probed every claim before making it, which is where this round came from: a centred fall-off for a glow, solid=True because density spaces the passes rather than filling them, a silhouette of its own for a round tip, a clean edge that stops insetting at the canvas frame, rehearsals numbered apart from the painting's looks, cost_line saying why a number is large, and smudge taking the boundary it is meant to run along.

The worked examples

paintings/ holds the paintings those sessions made, and each is an end-to-end worked example rather than a gallery: the numbered pass scripts that built it, the prelude.py of helpers and mixtures beside them, NOTES.md in the painter's own words, and the finished PNG and time-lapse. The scripts re-run from a fresh session at the same seed and reproduce the export byte for byte, so the order a painting was made in is readable rather than reconstructed — which is the one thing the guide cannot teach abstractly, and the thing a first-time painter is least sure of.

PAINTER.md deliberately does not point here, and that is the trade: a worked example names a subject, a named subject leaks, and six of six fresh sessions once painted a noun the guide had merely listed. A painter who goes looking finds these; one who only reads the guide is not handed a picture to paint. If you are about to run the measurement protocol in LESSONS.md, do not read them first.

PAINTINGS.md is the same two paintings read from the outside rather than from the painter's seat: what they cost, what failed, and how good they actually are.

Where the rest of it is written down

Four documents sit behind this one. PAINTER.md is the guide a painter reads and the project's actual deliverable, and REFERENCE.md is every fact on one page beside it — units, defaults, what each argument does — for looking up rather than reading. LESSONS.md is what six measured painting runs and an adversarial review left behind: the method, the engine decisions that are load-bearing, the traps, and what is still open; read it before changing the engine or the guide. SUGGESTIONS.md is the request list from the painting sessions, every item of which is now done, and it says what each one became.

Licence

MIT — see LICENSE.

The optional Mixbox pigment model gives better mixing than the built-in one, but its reference implementation is CC BY-NC. It is therefore an opt-in extra (pip install "easel-paint[mixbox]"), not a dependency — check that its licence suits your use before enabling it.