Gaea Media Server

Why a media server

Chat is one shape. Media is many.

Local LLM servers made open models easy to use, because text goes in and text comes out. Generative media doesn't fit that pipe. A picture needs a size, a seed, and maybe a mask and three references. A voice streams out while it is still being spoken. A video has a frame rate, a soundtrack, and a label saying a model made it.

Gaea takes that on so your app doesn't have to. A job names an operation and a quality tier. Gaea picks the model, fits it into memory, runs it, and hands back files that say what made them.

Chat Messagestext A replytext

image.edit Prompt and two reference imagessized to suit the model 1024 × 1024 PNGseed, lane and licence attached

audio.speak Text and a voiceor a reference to clone Speech, streamedchunks that rebuild the WAV

audio.generate "Rain on a tin roof", 6 swith a negative prompt 48 kHz audioexact to the frame

video.generate Prompt, first and last framesterms accepted by the user MP4 with soundlabelled AI-generated

Three profiles, one engine

Everything an app can ask for, by operation.

Gaea serves the Media Engine Spec's Image, Audio and Video profiles. Apps ask for an operation. Plugins bring the models behind it, and a new model reaches every app without an app update.

Image

MES Image

Generate, edit with references, repair and take apart.

image.generateText to imageZ-Image Turbo · FLUX.2 Klein 4B · Qwen-Image Flash · Boogu-Image · Ming-Image
image.editEdit from one or more reference imagesFLUX.2 Klein 4B · Qwen-Image-Edit 2511 · Boogu-Image
image.fill · image.upscale · image.restoreInpaint, enlarge and repairForge processing pack
image.matte · image.segmentCut out a subject, or select by pointsForge processing pack
image.decomposeSplit a design into layers at its own sizeMing-Image Design-Layer

Audio

MES Audio

Voices, edits, sound effects and whole songs.

audio.speakText to speech, streamed as it's spokenKokoro-82M, 28 voices · Qwen3-TTS 1.7B, voices, emotions and cloning
audio.design_voiceDescribe a voice and get oneQwen3-TTS Design
audio.edit_speechChange emotion or style, add a laugh or a sigh, denoise, trimStep-Audio-EditX
audio.generateSound effects from 0.5 to 30 s at 48 kHzMOSS-SoundEffect v2.0
audio.composeSongs and instrumentals from 10 s to 10 min, in 50 languagesACE-Step 1.5

Video

MES Video

Clips with sound, and the tools to finish them.

video.generateText or keyframes to video, with soundLTX 2.5, four memory tiers
video.extend · video.retakeContinue a clip, or redo a span of itLTX 2.5
video.upscale×2 or ×4, up to 1920 × 1080SeedVR2-3B
video.interpolate2× to 8× the frames up to 4K, keeping the soundRIFE 4.25

How it works

The host keeps order. The plugins do the generating.

Gaea runs the front door, the queue, memory across processes and each plugin's life. Every model family runs as a plugin in a process of its own, on the GPU your Mac already has.

Your apps

Your own appMESClientKit or plain HTTP

Gaea Labevery server as one engine

Pixel Agency appsDesktop and friends

Scriptscurl, CI, agents

MES · :17630

Gaeathe host

Queueone GPU job at a time

Memory governorbudget, lanes, eviction

Plugin managerinstall, update, roll back

Trust listsignatures, never URLs

Console, menu bar and the gaea CLI · gaea.sqlite

GPP · one process each

Plugins

z-image

flux2-klein

kokoro

ltx

MLX on Metal · unified memory

Add a plugin by its repo
gaea plugins add <github-url> fetches the newest release. Gaea checks its Developer ID signature and notarization against your trust list.
Its models appear everywhere
Every app sees the new operations and models on its next look. Models that need a licence accepted start switched off.
An app sends a job
It names an operation, inputs and a quality tier: draft, standard or best. Options a model adds travel as x- extensions.
Gaea makes room
It admits the job against the memory budget, stops idle workers if it must, and loads the largest lane that fits.
The plugin generates
In its own process, streaming progress and audio chunks as it goes. A crash costs one worker, and the job runs once more.
The result says what ran
Files with the model, licence, effective parameters and warnings. No media outlives its job.

What Gaea handles

The hard parts of running models, done once.

Memory that never surprises

Gaea knows a model's footprint before it loads it. Jobs are admitted against a budget of three quarters of your Mac's memory, and Gaea learns each model's real peak from its runs.

Lanes: bf16, int8, int4

When the default doesn't fit, Gaea loads the largest lane that does and adds a warning. Ask for a lane with x-gaea.lane and you get exactly that lane, or a clear refusal.

Trust from the signature

Plugins are notarized Developer ID bundles, checked at install and every launch. A release from a team you haven't trusted waits for your approval.

One process per model

A crashing model takes down its own worker, never the server. Gaea records the crash with macOS's report, reruns the job once, and quarantines a plugin that crashes three times in ten minutes.

Updates with a way back

Gaea and its plugins update from GitHub releases, verified before they install. Updates wait for running jobs to finish, and the previous version is kept for rollback.

Honest results

Every result names its model and licence, the parameters that took effect, and any warnings. Video carries an AI-generated label.

LoRA adapters by URL

gaea adapters add <hf-url> --model <id> registers a LoRA from Hugging Face, pinned to its commit and sha256.

Keeps nothing

Gaea is stateless and safe to restart. Inputs and outputs live only as long as the job, so libraries and projects stay in your apps.

For home and small offices

One Mac can serve the whole office.

Put a Mac Studio or Mac mini in the corner, add the plugins you need, and every app on the network can generate with them.

A menu-bar app, or headlessOpen Gaea.app, or run --headless from a launch agent on a Mac no one sits at.
Apps find it on their ownGaea advertises itself over Bonjour. It serves only its own Mac by default. --paired opens it to your network over TLS, and each device pairs once.
Several servers, one poolGive each Mac its own area, such as image on one and video on another. Apps pick a server for each job.
Only you configure itApps can use Gaea but never change it. Plugins, models and trust are set in its console or with gaea on that Mac.

Memory ledgerMac Studio · 128 GB

z-image-turbo · int8warm
kokoro-82mstreaming
ltx · standard64running
flux2-klein-4b · bf16idle, stopped first

Illustration. Gaea charges each worker what it holds, stops idle ones least recently used first when a job needs room, and refuses with no_capacity only when nothing more can be freed.

For developers

One API for every model.

Gaea speaks MES, an open HTTP spec for media engines. Write against operations once, and new plugins show up in your app without a code change.

# Ask any Gaea for an image
curl -s http://127.0.0.1:17630/mes/v0/jobs \
  -H 'MES-Version: 0.1' -H 'Content-Type: application/json' \
  -d '{"operation": "image.generate",
       "inputs": {"prompt": "a lighthouse on a cliff at dusk"},
       "params": {"quality": "best", "seed": 42,
                  "size": {"width": 1024, "height": 1024}}}'
# → {"id": "…"}

# Wait for it, then fetch result.outputs[].href
curl -s 'http://127.0.0.1:17630/mes/v0/jobs/<id>?wait=10' \
  -H 'MES-Version: 0.1'

# What can this server do right now?
curl -s http://127.0.0.1:17630/mes/v0/operations -H 'MES-Version: 0.1'

# Add models from GitHub
gaea plugins add https://github.com/xocialize-code/gaea-plugin-z-image
gaea plugins add https://github.com/xocialize-code/gaea-plugin-kokoro

# See what's offered, and choose a default lane
gaea models
gaea models lane z-image-turbo int8

# Keep weights on a fast volume
gaea models root /Volumes/Models

# Watch memory, workers and crashes
gaea ledger
gaea status

# Stay current, with a way back
gaea plugins check
gaea update --check

# 1. Start from the template
gh repo create my-gaea-plugin --template xocialize-code/gaea-plugin-demo --clone

# 2. Wrap your MLX model with GaeaPluginKit and describe it in
#    manifest.json: operations, quality tiers, lanes, licence

# 3. Sign with your own Developer ID, notarize, tag a release
git tag v0.1.0 && git push origin v0.1.0

# 4. Any operator can install it. The first release from
#    a new team waits for their approval.
gaea plugins add https://github.com/you/my-gaea-plugin

See it

A console for you. Any app for everyone else.

Gaea Lab generating a lighthouse at dusk. The model list shows Z-Image Turbo, FLUX.2 Klein 4B and Qwen-Image Flash with their licences, and a LoRA panel on the right. — **Gaea Lab**, the reference app, pools two Gaea servers into one engine. Each model shows its licence and memory size, and this image ran on Qwen-Image Flash with a LoRA.

Gaea's Plugins page showing the demo plugin with an update available, its signature and notarization, and the trusted teams list. — **The console's Plugins page.** Every update shows its signature, notarization and sha256 before you take it, and the trust list sits right below.

The plugins

Open models, ready to install.

Each one is a repo on GitHub with signed releases. Most come in several lanes, so a Mac with less memory runs a lighter quantization of the same model.

Plugin	Profile	What it does	Licence
Z-ImageTurbo and Base	Image	Fast text to image, and edits`int4 · int8 · bf16`	Apache-2.0
FLUX.2 Klein4B and 4B Base	Image	Generate, and edit from several references`int4 · int8 · bf16`	Apache-2.0
Qwen-ImageEdit 2511 and Flash	Image	Multi-reference editing, and fast generation	Per model, NVIDIA Open Model License for Flash
Boogu-Image0.1 Base and Turbo	Image	Generate, and one-image edit`int4 · int8 · bf16`	Apache-2.0
Ming-Image0.1 Design and Design-Layer	Image	Posters and layouts, transparent backgrounds, layers	MIT
Forgeprocessing pack	Image	Inpaint, upscale, matte, segment and restore in one worker	Per model
Prompt EnhanceQwen3.5 4B and 9B	Image	Rewrites short prompts where a quality profile asks	Per model
Kokoro82M	Audio	Text to speech with 28 English voices, streamed	Apache-2.0
Qwen3-TTS1.7B	Audio	Voices with emotion, cloning from a reference, voice design	Apache-2.0
Step-Audio-EditX	Audio	Edit speech: emotion, style, laughs and sighs, denoise, trim`int8 · bf16`	Apache-2.0
MOSS-SoundEffectv2.0	Audio	Sound effects from 0.5 to 30 s at 48 kHz`int4 · bf16`	Apache-2.0
ACE-Step1.5	Audio	Songs and instrumentals, covers, repaints and extensions`int4 · int8 · bf16`	MIT
LTX2.5	Video	Generate, extend and retake video with sound, in four memory tiers	LTX-2 Community
SeedVR23B	Video	Upscale ×2 or ×4, live action and drawn	Apache-2.0 weights
RIFE4.25	Video	Frame interpolation from 2× to 8×, up to 4K	MIT

Bring your own model: the plugin template starts you with signing, a self-check and release CI. Adapters are not plugins; add a LoRA with gaea adapters add.

Generate image, sound and video on your own Mac.