Download

For Emberite AI 0.42. The same guide is in the app's Help tab.

Emberite AI · Admin's Guide

The admin account runs the site: it sets up the GPU servers, chat, web search and the API for other apps. It doesn't chat or make images and videos. For that, sign in with an ordinary account.

First steps

  1. Sign in as admin. The initial password is emberiterising.
  2. Change the password at the bottom of the Admin tab. Until you do, a warning is shown there.
  3. Add more PCs as secondary servers if you have them, and set a Brave or Tavily key for web search (click Save).
  4. Check Chat models and Image and video models: after a Windows install, the models for this PC's graphics card are already downloading there.
  5. Check the Servers dialog: every server should show as up.
  6. People create their own accounts on the sign-in page (Create one). Each user only sees their own chats, tasks and files.

What the admin sees

  • Admin: the chat models, the site settings and the admin password.
  • Help: this guide. Ordinary users get the User's Guide instead.
  • Servers, Settings (theme) and Sign out at the top right. On a phone they're in the ☰ menu.

The Admin tab

  • The header shows the app's version, the LLM API address (http://<this site>/v1) with the number of requests since the app started, and the number of user accounts.
  • Yellow notes under it: while the admin password is still the initial one, and while there's no Brave or Tavily key (see Web search; without one the chat can't look anything up on the web).
  • Settings are grouped: Chat, Web search, LLM API, Code tools, ComfyUI and Files. Under each box is what it's for and its default. A changed box is outlined until you save.
  • Save checks every value and applies them straight away, with no restart. They're stored in the database. A bad value is outlined in red with the reason, and nothing is saved until it's fixed.
  • An empty box means the default. Secret values (API keys) are never shown again: the box says •••• 1234 (unchanged). Leave it empty to keep the key, or type a new one.
  • Reload workflow files reads the image and video models' workflow files again (see Image and video models).
  • Environment (read-only): EMBERITE_DATA_DIR (where the database, inputs and outputs are) and EMBERITE_COOKIE_SECURE (true when the site is served over HTTPS). These two are set where the app is started and need a restart. On the Windows install they're set from C:\ProgramData\Emberite AI, see Windows install.
  • Admin password: enter the current one and the new one twice.

Servers

The Servers panel at the top of the Admin tab lists every PC that does work: this PC (the primary, with the website), the secondary servers that joined it, and any ComfyUI or Ollama added by address.

Adding a PC

  1. On the new PC, run the Emberite AI setup and choose Secondary.
  2. Enter this primary's address and join key, both shown at the top of the panel (Copy key copies the key).
  3. The setup installs Ollama and ComfyUI there, and the PC joins on its own. It appears in the list within a few seconds, then gets the same chat models and image and video model files as this PC, in the background.

A secondary has no website: its address shows a status page (connected or not, its graphics card, its downloads). People keep using the primary's website, and their work is spread over all the servers.

The list

  • Status: online, or offline with when it was last heard from. Kept in step: how many of the image and video model files it has, its downloads, and its chat models. Files that are public (the Default models, files added by link) download from the internet; files you uploaded are copied from this PC.
  • Chat and Images: whether a server is used for chat and for images and video. Chat uses the servers in the list's order: the first one that's up (↑ ↓ change the order).
  • Rename: the name shown everywhere (tasks, the Servers dialog, users' Settings).
  • Remove: it stops getting work. A secondary that's still running joins again by itself, so click New key first (PCs that joined keep working; only new ones need the new key), or uninstall it on that PC.
  • A yellow note says when a PC also has another Ollama or ComfyUI (installed separately): uninstall that one if the disk gets full.
  • Add by address…: a ComfyUI and/or Ollama that doesn't run Emberite AI (for example a server set up before 0.39). It's used, but not kept in step or upgraded.

Upgrading every server

  1. Under the list, Upload a setup… and pick the new Emberite-AI-Setup-<version>.exe (keep its name: the version comes from it).
  2. Click Upgrade all servers. Each secondary downloads it from this PC and installs it silently; its status shows upgrading…, then upgraded.
  3. When they're done (or after 30 minutes, for ones that are off), this PC upgrades itself. The website is unavailable for a minute or two while it does. Settings, accounts and models are kept everywhere.

A secondary that joins later, or was off during the upgrade, is brought to the same version when it next connects.

Chat models

The Chat models panel at the top of the Admin tab lists the Ollama models on each chat server: one row per model, one column per server.

  • ✓ 6.6 GB: installed there. —: not on that server. ?: that server hasn't answered yet.
  • Pull: type a model name from ollama.com/library (for example qwen3.5:9b, or click a suggestion) and click Pull. It downloads onto every chat server in the background, with the progress shown per server. A server that's offline shows waiting for the server and gets it when it's back. A download that was cut off carries on where it stopped. A name that doesn't exist fails with the reason.
  • Pull to all on a row puts the model on the servers that don't have it, for example a chat server you added later.
  • Use for Chat makes that model the Chat model (the default for new chats and for the LLM API's default).
  • Remove deletes the model from every chat server straight away.

Users can pick any model that's on every chat server, at the top of the Chat tab. A model on only some servers isn't offered, because chat moves to the next server when one is down. Pick models that fit each server's GPU memory: a model that doesn't fit runs partly on the CPU and is much slower.

Models the Windows setup chose: the setup looks at the graphics card's memory and downloads, in the background, qwen3.8:27b for 24 GB or more; qwen3.5:9b and then qwen3.6:35b-a3b for 8 to 23 GB; qwen3.5:4b below that. The Chat model becomes the first one that's ready, and then the setup's choice (qwen3.6:35b-a3b in the middle case) when that's ready, unless you changed the Chat model meanwhile. The Chat tab shows the download progress to everyone.

Image and video models

An image or video model is a ComfyUI workflow plus the files it needs (checkpoints, diffusion models, LoRAs, text encoders, VAEs, upscalers). Emberite AI ships two: the image model Default (Qwen Image Edit 2511, about 31 GB of files) and the video model Default (LTX 2.3, about 43 GB). You can add more. Users pick their image and video model in their Settings; a model's other names (also called) work there too, and in task files.

The panel

  • One box per model, with its files: the folder each goes in, and per ComfyUI server ✓ (there), — (missing) or ? (that server didn't say). ✓ ready means every file is on at least one server; tasks go to servers that have the files first.
  • Downloads: on a Windows install the Default models' files download in the background into this PC's ComfyUI, one at a time, each checked against its SHA-256. A cut-off download carries on where it stopped, also after a restart. Without enough disk space it waits and says so. Until a model is ready, users see its progress on the Image, Video and E2E tabs and can't generate with it.
  • Upload… sends a missing file from your computer, under the name the workflow expects (whatever the file on your computer is called). Big files go up in pieces with progress; if the connection drops, pick the same file again (Carry on…) and it continues where it stopped.
  • Link… downloads a missing file from a link, optionally checked against a SHA-256 you paste. A Hugging Face page link (…/blob/main/…) is turned into its download link. Files that need a sign-in (gated) fail with a note: download them yourself and upload them.
  • Add a file to (under the models) adds any file to a folder, for example a LoRA you then add in a model's settings.
  • Remove on a file deletes it from this PC's ComfyUI. The Default models' files can't be removed.
  • Files are added to this PC's ComfyUI (models folder shown at the top). Secondary servers then get them too (see Servers); for a server added by address, copy the files there yourself.

Adding a model

  1. In ComfyUI, open the workflow and use Workflow → Export (API). (The normal Save format isn't accepted.)
  2. Click Add model… and pick that .json file. The app finds the parts it fills in (prompts, input image, size, length, frame rate, the save node) and lists the files it needs. Files from Emberite AI may already carry a name, type and other names.
  3. Check the Type (Image or Video), the Name and Also called, and click Add model. Then add its missing files.

What an image workflow needs: a Save Image fed by VAE Decode ← KSampler, whose positive and negative prompts come from text encoders (Qwen image edit or CLIP Text Encode); optional Load Image nodes for the input and reference images. A video workflow needs: Load Image, Save Video, and primitive nodes titled Prompt, Width, Height, Duration and Frame Rate; optionally a Boolean titled prompt enhance and a LoRA titled Style LoRA (users then get a strength box). If something is missing, Add model says what.

Settings (the editor)

  • Settings on a model lists what can be changed, node by node: the model files, LoRAs and their strengths, samplers, sigmas, CFG and the like (not the parts the app fills in per task). Save applies to new tasks.
  • Remove on a LoRA takes it out of the workflow. Add a LoRA: the file (in loras), its strength and where it goes (after the model loader or another LoRA).
  • The Default models can be changed too; Reset puts back the settings they shipped with.
  • Workflow downloads the model's workflow (with your changes), to keep or to add on another Emberite AI.
  • Remove on an added model removes it (its files stay). Its users go back to Default.

Chat

  • Where chat runs: the servers with Chat ticked in Servers, in the list's order. The first one that's up is used; if it goes down the next one is, and chat moves back when the first is up again. With no chat server, chat is off (the Chat tab says so, and the LLM API answers 503).
  • Chat model: the default model for new chats and for the LLM API (default qwen3.8:27b). Users can pick another in the Chat tab. Get models onto the chat servers with Chat models; Use for Chat there sets this box.
  • Chat idle minutes: how long without chat, or LLM API requests, before the model is unloaded and the server goes back to images and videos (default 30).
  • Chat servers at most: how many servers chat may take when it gets busy (default 2). Each chat server answers one request at a time (Ollama runs with one request at a time, so the whole context fits in graphics memory). When requests have waited about 10 seconds, the next reachable server with Chat ticked in Servers, in the list's order, is reserved too: its tasks move away, the chat model loads, and requests go to whichever chat server is free. An added server goes back to images and videos after 5 minutes without requests (or with the first one, or Release now). 1: chat never takes more than one server.
  • Chat context: how many tokens the chat model reads at once (Ollama's context length), on every chat server. auto (the default) picks by each server's graphics memory: under 8 GB 16K, 8 to 23 GB 32K, 24 to 31 GB 64K, 32 GB or more 128K. Or enter a number such as 64K or 40960. A larger context needs more graphics memory, and a model that doesn't fit spills into normal memory and gets much slower. Coding tools and agents send long instructions: Claude Code and Hermes Agent want 64K or more, OpenHands at least 32K. Saving a new value restarts Ollama on each PC (the primary at once, secondaries within seconds), which unloads the chat model; the Servers list shows each server's context. It applies to Emberite AI's own Ollama on Windows; for a server added by address, set OLLAMA_CONTEXT_LENGTH there yourself.
  • Chat API key: only if your Ollama (or the proxy in front of it) needs a bearer token.

How chat shares a GPU

A ComfyUI server that's also a chat server is reserved for chat while people chat:

  1. The first message reserves it. Its image/video tasks are interrupted and moved to the other servers, or they wait (shown as waiting) if there's nowhere to go. ComfyUI is asked to unload its models, and the chat model loads.
  2. New tasks avoid it while it's reserved.
  3. After the idle minutes (or Release now in the Servers dialog), every chat model loaded there is unloaded. Waiting tasks start again, and queued tasks on the other servers move back when that finishes them sooner.

A request that's still running is never cut off by the idle timer.

LLM API for other apps

The site has an API at http://<this site>/v1 (shown in the Admin tab header) on the same port as the site. It speaks OpenAI's Chat Completions and Responses APIs and Anthropic's Messages API (through Ollama), so OpenCode, Claude Code, Codex CLI, Cline, OpenHands, Hermes Agent and similar tools connect to it directly. Requests go to the chat servers exactly like the Chat tab, with the same reservation and idle time. The model default means the Chat model setting, and so do the cloud model names coding tools send on their own (claude-…, gpt-…, o3, codex-…); Ollama names such as gpt-oss:20b are kept.

  • LLM API key: when set, apps must send it as their API key (Authorization: Bearer …). Empty: no key needed, so anyone who can reach the site can use the model. Set one if the site is reachable beyond people you trust. The key is never passed on to Ollama.

Code and Agent tabs

The users' Code tab gives setup steps for OpenCode, Claude Code, Codex CLI, Cline and GitHub Copilot, and the Agent tab for Hermes Agent and OpenHands, with this site's address, the chat model, the LLM API key and the chat context filled in. Every tool talks to this site directly: Claude Code → Emberite AI → Ollama. There's nothing to set up for them beyond the LLM API and Chat context.

The User's Guide tells users what to expect from Claude Code with a local model (focused tasks work best, no Anthropic server tools). Pick a chat model that's good at tool calls, such as a recent Qwen coder model, and give it 64K of context or more.

ComfyUI

  • Which servers make images and videos: the ones with Images ticked in Servers. Users on Auto get whichever server should finish their task soonest; this PC is the default.
  • Removing a server doesn't delete its tasks. Its waiting tasks can be moved with Switch tasks.
  • Poll interval (s): how often the app checks ComfyUI for progress (default 2).

Files

  • Max image upload (MB): the largest input image (default 25).
  • Max video upload (MB): the largest video a user can give as a video's input, to use its last frame (default 1024).
  • ffmpeg command: used to read a video's last frame on the server when the browser can't (for example H.265). Default ffmpeg, which must then be on the PATH. Optional.

The Servers dialog

Click Servers at the top right. The dot on the button is green when every server is up and idle, amber when work is queued, blue when a server is reserved for chat, and red when one can't be reached.

  • One row per ComfyUI server: its address, whether it's up, offline or in use for chat, the running and waiting tasks (everyone's), and the estimated work queued.
  • When a server is reserved for chat, a line says which one, the model, and when it frees up. Release now hands it back to images and videos straight away.
  • Typical generation time: the median time for each kind of task on each server, learned from finished tasks (×N is how many it's based on).

Settings

For the admin account, Settings (top right) is just the Theme: Dark or Daylight. The dialog also shows the app's version. Each user sets their own theme and preferred server in their own Settings.

Windows install

When Emberite AI was installed with Emberite-AI-Setup-<version>.exe:

The setup

  • Primary or secondary page: install one primary (the PC with the website), then choose Secondary on every other PC, with the primary's address and join key from Servers. A secondary always installs ComfyUI (it copies the primary), gets its chat models from the primary, and shows only a status page. For a silent install: Emberite-AI-Setup-<version>.exe /S /ROLE=secondary /PRIMARY=192.168.1.20:8088 /JOINKEY=<key>.
  • Graphics card and models page: it shows the graphics card it found and its memory, and which chat models that means (see Chat models). If the memory is wrong, type the right number. That's common for integrated graphics with shared memory, which report what the BIOS sets aside. The page updates as you type.
  • Chat models folder and Image and video models folder: by default inside the install folder (C:\Program Files\Emberite AI\ollama\models and …\ComfyUI\models, or on D: if you installed on D:). Pick other folders (for example on a bigger drive) with Browse. The image and video models need about 74 GB.
  • It downloads Ollama 0.40.5 (about 1.5 GB, plus about 250 MB for AMD graphics) and checks it against the release's SHA-256 before installing it in C:\Program Files\Emberite AI\ollama.
  • With 8 GB of GPU memory or more it also downloads ComfyUI 0.41.0 (the portable build for NVIDIA, AMD or Intel graphics, about 2 GB, SHA-256 checked) and unpacks it into C:\Program Files\Emberite AI\ComfyUI (it needs about 10 GB free while installing). Below 8 GB there's no ComfyUI: chat only.
  • Emberite AI runs its own Ollama and ComfyUI. If another program already uses port 11434 or 8188 (for example an Ollama or ComfyUI installed separately), its own get the next free port (11435, 8189…) and the app is set up to use those.
  • Without internet, or if a check fails, Emberite AI is still installed; run the setup again to retry.

Running

  • Three scheduled tasks start with Windows (no one needs to be signed in) and start again if they stop: Emberite AI (the site), Emberite AI Ollama and Emberite AI ComfyUI.
  • Ollama (port 11434) and ComfyUI (port 8188) listen on the whole network, and the firewall is open for them on private and domain networks, so other Emberite servers can use this PC. Neither has a password: anyone on your network can use them.
  • Start menu → Emberite AI has: open it, Restart and Stop (the site, Ollama and ComfyUI), Log file, Ollama log file, ComfyUI log file, Settings (port), Data folder and Uninstall.
  • The data (database, inputs, outputs, added models) is in C:\ProgramData\Emberite AI\data, ComfyUI's own input and output in …\comfyui, and the logs in …\logs (emberite-ai.log, ollama.log, comfyui.log). They are kept on upgrade and uninstall. Back up that folder to keep everything.
  • C:\ProgramData\Emberite AI\emberite-ai.ini holds the port, the address to listen on (host, 0.0.0.0 for every network), cookie_secure, and Ollama's and ComfyUI's models folders and ports. Use Restart after changing it. If you change a port, the Windows firewall rule still has the old one: run the setup again and it fixes it.
  • From SmartEmber AI (the name before 0.38): run the Emberite AI setup. It moves everything over: the data, settings and accounts (to C:\ProgramData\Emberite AI), the install folder with Ollama, ComfyUI and the downloaded models (to …\Emberite AI next to the old one), and it replaces the old scheduled tasks, firewall rules, Start menu entries and the entry in Apps. Nothing downloads again. If the admin password was still smartemberrising, it becomes emberiterising.
  • To upgrade, run a newer setup. Settings, accounts, files and models are kept, and Ollama and ComfyUI aren't downloaded again unless their version changed.
  • Uninstall removes the site, Ollama and ComfyUI. The downloaded models are kept unless you tick Also delete the chat models or Also delete the image and video models.

On a phone

Tap ☰ (top left) for Admin, Help, Servers, Settings and Sign out. The settings are one per row, with the help under each box.