Easel

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.

You have to be able to see images. Every pass of the method ends by looking at what it did — s.look(), the rehearsal, the post-pass check — and none of that reaches a painter who cannot take an image back. A model without image input can drive the whole API and will get a picture out of the far end, but it cannot tell whether any of it landed; one painter in the 0.5.0 cohort said exactly that of its own finished painting. Vision is a requirement here, not a feature.

Read the guide before you paint. The engine is only the brush; PAINTER.md is the method, and six measured runs say the method is the half that matters. It ships inside the package, so there is nothing to go and find: easel guide prints its first page — the whole workflow, about 1,400 words — and easel guide --full prints the rest.

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 structures.stroke([(0.2, 0.6), (0.6, 0.55), (0.9, 0.62)], "bristle", "yellow_ochre")s.export("painting.png")

Paintings, made this way

[Inside a car wash seen from the driver's seat: a magenta foam arch overhead, a bloom of white light down the tunnel, a red stop light, and a foam-covered side brush swinging in from the right, past a steering wheel and rear-view mirror]

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 it was painted from are beside it.

[A laundromat at night seen from the sidewalk opposite: a wide lit shopfront window set in a dark building, a row of washing machines with round doors inside it, one person sitting alone at the right-hand end, and the window's light broken into streaks down a wet road in the foreground]

A laundromat at night, from across the street — 286 strokes of a 300 budget, 1024×768 linen, no reference photograph. The fourteen pass scripts it was painted from are beside it.

[A grey heron seen from high above in a flooded parking lot at dawn, the water filling nearly the whole frame: a near-black band of trees across the top with a warm yellow reflection under it, the bird standing large and high in the picture with a white head and a long ochre bill and its own pale reflection directly below it, and painted stall lines running away across the water in the foreground]

A heron in a flooded parking lot at dawn — 253 strokes of a 320 budget, 1024×768 rough canvas, no reference photograph. The same subject was painted twice by the same painter to separate two kinds of fault; PAINTINGS.md has both attempts side by side.

Every one of these was painted by a language model working from PAINTER.md, one call to this API at a time, with no human hand on the canvas and nothing traced. Every stroke is in the log, every time-lapse was 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: have you decided what to paint? If you have, these are yours to study and the scripts beside them are the best thing here. If you have not, do not look — a worked example names a subject and a named subject chooses for you, as six of six fresh sessions once demonstrated by painting a noun the guide had merely listed in passing. Decide first, then look.

The brushes themselves

[Brush sampler: every brush at three sizes and three pressure profiles, on smooth, linen and rough canvas]

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.

[Shape sampler: five ways to build a mass, each filled along four sweep directions, with the bounding box in the last column]

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 APIpip install "easel-paint[mcp]"   # and the MCP server

from easel import Session, blob, cell      # the import the documents teachimport easel_paint                         # the same engine under the installed name

The distribution is easel-paint, never easel. PyPI carries an unrelated package called easel, so pip install easel gets somebody else's. Both import names work here — easel_paint re-exports easel — and every example is written with the short one.

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 ten warm-up exercises come before anything else.

If you have already installed the package you do not need that link. All six documents ship inside the wheel: python -m easel guide prints the first hour, --full prints the rest, and easel.docs.read() hands you any of them as text. The sixth is DIAGNOSIS.md, an index of symptoms rather than a document to read — you have rehearsed a pass, it is wrong, and easel diagnose concentric rings prints the passage that measured it.

It is one file of five, and every rule in the five is stated once, in the file it belongs to, and linked from everywhere else. PAINTER.md is held to a word budget by the test suite so that it stays the file a painter can hold in their head; the numbers it quotes are measured in CALIBRATION.md, and what painters reported about their own sessions is kept there too rather than beside the rules.

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 1,500 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 300easel run painting.easel first_pass.pyeasel run painting.easel first_pass.py --thumbnail  # its masses, flat and small, unpaintedeasel run painting.easel first_pass.py --rehearse   # against a copy, committing nothingeasel run painting.easel p2_sea.py p3_rocks.py --rehearse   # ...both passes, one copyeasel run painting.easel sky_a.py sky_b.py --alternatives   # ...two versions, side by sideeasel run painting.easel p4_tower.py --check         # the post-pass check over the whole paintingeasel look painting.easel --grideasel look painting.easel --valueseasel look painting.easel --region D4 --fine --reference ref.jpgeasel mark painting.easel top_l 0.335 0.315easel compare painting.easel ref.jpgeasel prepare painting.easel ref.jpg --level coarseeasel undo painting.easel 3easel export painting.easel painting.pngeasel timelapse painting.easel painting.gif --every 3 --scale 240easel brusheseasel guide                                          # the method, in about 1,400 wordseasel guide --full                                   # all of PAINTER.mdeasel guide --painting                               # the reasons under the ruleseasel guide --recipes                                # the calls, in order, for a thingeasel guide --reference                              # units, defaults, every argumenteasel diagnose concentric rings                      # it is wrong: the passage that says whyeasel explain chisel-blank                           # it warned you: the measurement behind iteasel demo crosses a boundary                        # a recipe, painted beside its failure

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, thumbnail, preview, rehearse, compare and prepare return their PNG inline, as run does when it rehearses a pass, 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.

Twenty-one tools: the seventeen CLI verbs, plus preview, rehearse, cost and thumbnail — the four questions about marks that have 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; run(alternatives=[...]) tries versions of one, each on a copy of its own, and hands them back side by side; run(thumbnail=true) counts a pass and hands back its masses flat and small, before any of it is paid for. 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: to the pixel on one machine, and within one level of 255 on another, since a numpy build can round a few pixels differently. Nine pieces painted on Linux rebuilt to the pixel there, and on Windows within one level, in 12 to 2,508 pixels a piece. 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 exports.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, and the rebuild is laid by the engine installed. So is easel timelapse, since the file keeps no frames: the film is the log replayed, frame for frame the one the painting recorded. A file saved by an earlier release says so when it opens, if a fix since then lays some of its marks differently.

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.
• A brush running dry drags. Gated a pixel at a time, a starving brush still left dots at the weave's scale -- confetti, which reads as dirt. So as its load falls the gate reads the tooth along the stroke's travel, and a bristle comb's bristles run dry one by one; both are matched to the tooth's own distribution, so a load lays what it laid and only the shape changes, into streaks that run with the 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 with every change since.

Every one of those pieces was added or corrected because a painter actually used the guide against the engine and wrote down what it cost them — an engine change lands with a test and a measurement behind it, and a guide change stays a hypothesis until the next session paints against it and says so.

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 — 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.

Two different claims live here, and this page used to make the stronger one about every painting. The one that holds everywhere is that the log replays to the stroke: a stroke's randomness is drawn from (seed, stroke index), so a saved painting's log lays every mark again where it lay, and golden-image tests hold that. To the byte it holds on one machine: two builds of numpy can round a mixture differently in the seventh decimal, which moved 729 of one painting's 786,432 pixels by one level. It is laid by the engine installed, so a stroke an earlier version laid wrongly comes back fixed, and CHANGELOG.md names each such fix under its version. Whether the committed scripts rebuild the canvas is a separate question and the answer is per painting — seven of the thirty-four say Reproducible: not claimed in their own table in PAINTINGS.md, usually because a drawing pass was rewritten and re-run, or look scripts ran as passes between the painting ones, and a mark's texture is seeded from its place in the log. Each painting's table says which of the two it claims; the older ones that claim the stronger one were checked by sha256, and the seven newest carry their own painter's claim, not yet re-run here.

PAINTER.md points here with a condition attached, and the condition is the whole trade: if you chose your subject before opening this repository, these are yours; if you have not chosen, they will choose for you. A worked example names a subject and a named subject leaks — but a painter who decided first cannot be steered by a noun, so the cost of hiding these no longer has to be paid by the painter who would most benefit from them. If you are about to run the measurement protocol in LESSONS.md, that rule is the protocol, not a suggestion.

PAINTINGS.md is the same 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

The guide's own five files are in the table above. Two more sit behind them, for somebody working on the project rather than painting with it. 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, because it is also where the rules about how the guide may change are written down. SUGGESTIONS.md is the register of what twenty-five painters asked for after using the guide: what was wrong, and what was done about it, one pair of lines each. Everything on it is done — including the ten that were answered by measuring them and finding nothing to fix, which say so — except the round at the top, which is open and says so on every line. CHANGELOG.md is the same history cut by release rather than by painter: what changed in each version, which defaults moved, and what a script that leaves an argument off will paint differently after upgrading.

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.