Connected GPUs
—Awaiting serverAI Server.Private AI workspace
AI SERVER
Live
WORKSPACE / OVERVIEW
Dashboard
Monitor the remote Ubuntu AI appliance · Target: Not connected
Ubuntu appliance connectionWaiting for live hardware, models, services, and installed-software data from the configured server.
Your workspace
Connect to your server to see what is ready.
Tailscale remote access
Provided address · Connection not checked
Connect this device to your Tailscale network. This shortcut opens a new tab; your current server connection stays unchanged.
GPU memory
—Total across the GPU poolGPU power
—Combined GPU draw · excludes hostServer uptime
—Since the last restartLoaded model Awaiting server
Waiting for the Ubuntu appliance…
Backend—
Mode—
Context—
n_ctx / n_batch—
MTP Modules—
GPU Split—
Threads—
Status—
System resources
CPUAwaiting serverTemperature: —
—
RAMAwaiting server
—
System driveAwaiting server
—
OSAwaiting Ubuntu appliance
Uptime—
GPU FLEET
GPU activity Awaiting server
Live activity, power limits and supported LED controls for each GPU. Settings are saved for reboot.
Case Fan Hub
Your five case fans follow the hottest GPU. Choose a preset or adjust the curve below. CPU and chipset cooling stay automatic and separate.
Checking fan status…
Double-click UPDATE-AI-SERVER.bat to finish setup. The updater detects connected case fans and asks for your hub connector and BMC login once. No Ubuntu commands are needed.
GPU fan curve
Requested hub speedDrag points to shape the curve. Double-click the graph to add a point.
Control points
Model setup
Manage GPU allocation, load a saved profile or configure another model.
GPU allocation
Select free GPUs for your next model. Release an occupied GPU by stopping the model using it.
Checking GPU availability…
Loaded models
Checking loaded models…
Launch a model
Start with a saved profile, or choose a model and backend below.
Enter a model path or repository
3. Parameters
Backend reference ↗Select an option to add its line, then edit the value. You can also paste a working launch command.
@ lines are AI Server settings; export lines set runtime variables. Changes apply when you start the model.
Acceleration and option help
Search MTP, DSpark, DFlash, DFlash2, EAGLE3 or n-gram in the option picker. Choose a compatible helper when required. Options depend on the model and installed backend version.
Choose a model or load a saved profile to begin.
Stop all models
Stop every loaded language model. To release specific GPUs, use the controls above.
Installed models—
Library size—
Active nowNone
| Name | Location | Path / ID | Size | Last used | Loads | Actions |
|---|
Downloads & merges
Downloads and GGUF merges run in the background. Keep using the dashboard, or close your browser and return later.
Checking downloads…
Downloads can run concurrently; pause any you want to hold. Speed measures model bytes processed, not exact network usage. Active downloads and recent results are shown for the current dashboard service session; restarting the server/service clears this list.
Checking downloads…
Find a Model
Search Hugging Face or ModelScope to browse models.
Search Hugging Face or ModelScope, or paste a model URL/ID below.
Checking Windows Model Library…
Hugging Face: fast mode uses parallel files and Xet chunks. More connections can help a fast connection; reduce them if your disk or network slows down.
Downloads run in the background. Check the Download Queue tab or the Download Queue button at the top for progress, pause and cancel.
Auto detect understands Hugging Face, ModelScope and Ollama URLs. Direct HTTPS downloads reject local/private network destinations and unsafe redirects.
Hugging Face Best Practices
Read from this model's current Hugging Face model card when you inspect it.
Select a Hugging Face model to load its model-specific guidance.
DFlash2 Acceleration
Checking…Find target-specific DFlash draft models for the selected Hugging Face model.
Inspect a Hugging Face model to check DFlash2 compatibility.
Model Files
Choose a search result or inspect a URL/repository to view downloadable files.
Saved Model Profiles
| Name | Backend | Model | Owner | Actions |
|---|
Save the model editor as a profile
Saves the model, backend and parameters currently shown in Dashboard → Model setup. Load a profile there to edit it.
Understand the test. Review the model. Compare what it produces.
Run setup
Choose a model, review a test, then start.How model selection and harness tests work
Only the selected model receives requests. Browse and review tests before choosing Start. The default is selected in Coding workspace → Harnesses. Harness tests run the installed OpenCode, Pi or DeepSeek engine with a fresh, no-tools test profile and open its live session view here automatically. They do not use your normal project, extensions or file permissions. Their results are separate from the direct-model leaderboard. Standard external suites keep their own evaluator.
Generation & comparison settings Output budget, time limit, seed and labels
Choose a preset or enter 1–1,048,576 tokens. Maximum available uses the loaded context minus prompt space and a safety reserve, capped at 1,048,576 output tokens. If prompt tokenization is unavailable, a conservative estimate is recorded. Fixed budgets are also reduced when needed to fit known context. Backends may impose additional limits; a budget is a ceiling, not a required response length.
Large budgets can take hours and include thinking where applicable. Changes apply to new runs only. Keep the same model loaded and compare matching actual budgets, seeds and variant counts. These output controls apply to local v2 and one-shot creation tests; standard external suites, legacy tests and speed tests keep their own settings. Temperature is zero, but backend determinism is not guaranteed.
Review · nothing starts automatically
Review benchmark
All model and configuration details recorded for this run
Quick checks & standard tests
Start with a quick check, compare quality with the 32-test suite, and measure speed separately. Every selection opens an explanation; only Start benchmark sends requests.
More tests · standard suites, reasoning and long context
Build a custom capability run —
Published benchmark references
Integrated benchmarks run from the presets above. Only suites with a runnable card are integrated. Other links remain references.
What each benchmark measures
Local accuracy32 original tasks across eight capabilities; not an IQ test
Response speedEnd-to-end latency and effective output rate, including prompt processing
CodingCode reasoning locally; executable correctness in sandboxed EvalPlus
ReasoningGSM8K, BBH and MMLU-Pro metrics remain separate
WritingHuman-review rubrics; no automatic quality score
Long contextThree retrieval probes; labels estimate tokens from character count
Tool decisionsSimulated selection and planning, not real agent execution
ServingvLLM concurrent serving throughput; separate from quality
CADBench, GameDevBench & LiveCodeBench
Run pinned benchmark tasks on your selected loaded model. Each choice opens a review first; completed results join the matching leaderboard automatically.
Checking external runtimes…
Preparing all suites downloads about 5.2 GB of datasets plus the pinned Blender/Godot container. CPU rendering and code execution use isolated containers; your loaded models stay running. Preparation does not send benchmark prompts.
Seed selects the same task subset across models. External suites use one attempt per task. CADBench records the local judge configuration; judging a model with itself can bias scores. GameDevBench uses a bounded dashboard agent and upstream Godot validators. LiveCodeBench uses release_v6, the lite tests and one greedy Python solution. Subsets and different judges/agents/runtime limits remain separate protocols.
Visual projects & complete 3D games
Build editable Blender scenes and complete playable Godot games. Review the selected model, workload and judge before starting. Finished scores link back to the native projects, renders and actual gameplay recordings.
Checking visual runtimes…
Preparation downloads about 473 MB of briefs, references and Kenney art packs, plus Blender 4.2.10, Godot 4.6.2 and container layers. It runs fixed CPU engine checks without sending model prompts. Full game generation can take hours; start with one task.
These use a bounded local project agent and local vision judge. Scores are separate from published rankings and from runs with different judges, tasks or limits. Image/video briefs need a vision-capable candidate. CPU tools use 8 CPUs / 16 GiB; rendering and evaluation follow the generation limit. Project cap: 512 MiB; each saved file or ZIP: 128 MiB; evidence: 2 GiB per task. A judge reviewing itself can bias results.
Complete-project defaults: up to 8 hours per task, 192 agent turns and 32,768 output tokens per response. Choose Custom duration to enter minutes, or No time limit to remove the generation deadline. Time is shared across all turns and tool work; turn/output limits and individual tool/evaluation limits still apply. Reasoning and the answer share each response’s token cap. Review shows the selected limits before starting. These settings apply to new runs.
What do these advanced tracks test and save?
BlenderWorld: 223 official text, image and video briefs, hidden weighted criteria, eight adaptive renders and actual scene inspection. Saves native .blend files, GLBs, renders, animation previews when present and judge evidence.
GameCraft-Bench: 140 official complete-game briefs. The agent builds a Godot project and demonstrates it through real keyboard/mouse traces; the upstream replayer records MP4 gameplay. Saves the full project, videos, sampled frames, logs and rubric grades.
Expert Blender + Godot: five original hard missions: mechanical escape, suspended cargo physics, flooded ruins, 3D rescue tactics and vehicle delivery. Each requires at least four Blender-authored asset groups used in an interactive 3D game, plus a title, controls, objectives, failure, pause and restart. Gameplay, visual quality, asset integration and completeness receive separate grades.
Blender, Godot & coding challenges
Pick challenges, test the output, then add your ratings to a model ladder.
Start at Level 1, then move up. Each test requests a complete HTML, SVG, Blender Python or Godot 4 GDScript artifact in one response: no follow-up turns, tools or repair attempts. Levels describe task difficulty, not guaranteed model sizes.
Loading creation prompts…
Output budget and time limit are shared with Comparison settings above and apply to new runs only. No time limit waits until the response finishes or you cancel the job; backend limits can still apply. Task default keeps each card’s original token cap; select a larger budget or Maximum available for longer artifacts. Thinking may count toward the allowance. Seed and variant settings do not change these one-shot runs. Token-limited answers are marked incomplete.
How to compare and review safely
1. Select challenges and run them on your loaded model. 2. Open Inspect answers below and copy or download the full response. 3. Test it and rate every rubric item 0–4: absent, weak, partial, good, complete. 4. Save review & add to ladder. 5. Use Compare next model, load another model, then Review selected and Start benchmark. Reuse the same browser/Blender/Godot version, prompt set, budget and harness version; protocols keep differing runs separate. Each task has equal weight. Only your latest reviewed run per model/configuration is ranked, not its best attempt.
Generated code is untrusted. The dashboard never renders HTML/SVG or executes Python. Inspect the source first; use an isolated browser or disposable Blender/Godot environment without sensitive files or network access. Copy the reviewed contents into the named file format to test it. These results are human-reviewed and never enter the automatic accuracy leaderboard.
Which benchmarks should I use?
After each tuning change
Run the 8-task smoke test to catch obvious breakage, the 32-task local suite for a broader regression check, and response speed separately. Small local sets help you iterate; they do not establish general model superiority.
Stronger quality checks · available here
IFEval checks precise instruction following. HumanEval+ and MBPP+ execute Python tests in the existing sandbox. MMLU-Pro and reasoning suites provide additional named metrics. Keep exact evaluator versions, prompts, sampling and budgets fixed.
LM Evaluation Harness ↗ · EvalPlus ↗Worth adding for deeper comparisons
LiveCodeBench adds newer coding problems; record its release and date range. RULER tests several long-context abilities beyond a single retrieval question. LiveCodeBench is integrated above using its pinned Python checker. RULER remains an external reference; its link does not launch a run.
LiveCodeBench ↗ · RULER ↗Blender · geometry, rigging and rendered output
The Blender challenges below progress from a robot to an animated observatory. Review script errors, geometry, dimensions, materials, camera views and animation frames in the same Blender version. Save screenshots and test observations in your review notes. The CADBench runner above adds automatic renders and local vision-judge grading of the published criteria. It does not establish general animation quality, and local judge scores remain separate from the paper’s GPT-4o results.
BlenderLLM / CADBench methodology ↗Godot · working games and engine APIs
Five Godot 4 challenges below cover signals, breakout, platforming, navigation and tactical game rules. Test the generated script in a fresh project; rate actual gameplay, errors and resets. GameDevBench evaluates game-development tasks, and GameCraft-Bench evaluates complete playable projects with replayed interactions. GameDevBench runs above with upstream validators and the dashboard agent. GameCraft-Bench and the expert Blender + Godot track run above and save complete projects and recordings. Godot engine FPS benchmarks measure the engine, not the model’s coding ability.
GameDevBench ↗ · GameCraft-Bench ↗Coding · functions versus real repository changes
Use EvalPlus here for executable Python correctness. LiveCodeBench adds newer contest-style tasks. SWE-bench tests fixes in real repositories, so pin the agent, tools, container and test version as well as the model. These measure different skills from producing a polished app in one response. External SWE-bench is a reference, not an installed dashboard runner.
SWE-bench ↗Use your own real work as well
Keep a small fixed set of coding changes and documents representative of your workload. Use the same harness and grading rubric, and review failures. A model can improve on one test while becoming worse for the work you actually do.
Your model & configuration leaderboard
Automatically records the tested model and settings. Rank quality, coding accuracy or response speed separately; choose the same test protocol to compare tweaks.
| Rank | Model / configuration | Quantization | Context | Backend / GPUs | Result | Run / compare |
|---|---|---|---|---|---|---|
| Run Local leaderboard · 32 tests on each loaded model to compare real scores. | ||||||
Choose Compare on two rows to see setting differences. Missing historical metadata is shown as not recorded. Accuracy, speed, human ratings and distinct standard-suite metrics are never averaged into one score.
Original task accuracy is not an IQ score or a published benchmark score. Standard-suite metrics are exported separately. No scores are supplied until a model actually runs.
Your Blender, Godot & coding ladder
Your saved rubric ratings, separate from automatic accuracy. Matching prompts, output caps and execution protocols only; direct and coding-harness runs never mix. Latest reviewed run per model/configuration within the 100 recent visible runs. Unrated runs are not treated as zero.
| Rank | Model / configuration | Quant / context | Backend / GPUs | Rubric score | Tasks | Run / compare |
|---|---|---|---|---|---|---|
| Run a challenge, inspect the results, then save your review. | ||||||
Latest capability breakdown
Live benchmark activity
Choose any run to follow model output, tool calls and errors, or replay its saved activity.
Model output
Waiting for model output.
No reasoning text received.
Tool calls, runtime messages & errors
Waiting for benchmark activity.
Task prompt
You can navigate away and return; tasks continue on the server. Select an earlier run or task to review saved output. No time limit is the default; Cancel stops the run. Every new benchmark streams its activity here, including direct-model tests, external evaluators, visual project agents and DeepSeek Harness. Runs use Benchmarks/<run ID> under the server project folder. Full activity stays saved if you leave this page. Live capture adds some timing overhead, so captured runs use a distinct comparison protocol.
Benchmark history
Workspace preparation enables the DeepSeek live panel and organizes saved projects. Recent runs and their outcome appear below, newest first. Remove entry deletes its saved result, transcript, reviews and leaderboard entry after confirmation. Cancel active runs first. Exported copies and disk artifacts are retained.
| Time | User | Preset | Model | Backend | Status | Score / Result |
|---|
Model output, files and grading
Select a run to read model answers, agent turns and tool feedback, or inspect renders, gameplay, expected results and grading. DeepSeek Harness answers and exposed reasoning are available in its live / saved transcript.
Installed Coding Harnesses
Install only the harnesses you want. All three can coexist; one can be selected as the default for ai-code.
Model defaults & helpers
DeepSeek, Pi and OpenCode · global across every project
Loading ready models…
Choose from all loaded models. Auto uses the saved model when ready and falls back to the first ready model when it is unavailable. This also sets Auto for connected apps, including Open WebUI. Save below, then refresh DeepSeek to update its label. Models stay running.
Loading global delegation settings…
Choose all loaded models to discover helpers automatically as models change. Independent analysis, reviews and drafts can run across one, two or more helpers; the lead supplies context, applies edits and verifies results. No helper means the lead continues alone. Save, then start a new harness chat or reload its tools. These controls do not load or unload models. Actual delegation and speed depend on the lead model.
Checking automatic model connections…
Load a model on Dashboard, then open your app. Managed defaults follow AI Server → Auto automatically. Open WebUI launch buttons preselect Auto, and OpenCode, Pi and DeepSeek use it by default. Checks run about every 15 seconds. Existing sessions and explicit model choices are not rewritten; select Auto once or start a new session if needed. Pi refreshes on ai-pi launch. No app is restarted automatically.
DeepSeek says “settings are unavailable in this browser”?
This developer-preview version restricts Settings when opened through a remote address. AI Server manages the local provider and shared Harness process on the Ubuntu server, so close Settings and use the chat model picker. This message does not mean the model failed to load. Do not disable authentication or browser security to bypass it.
Harness changes are Administrator-only. Browser DeepSeek uses the shared /srv/ai/projects workspace and shared Harness history for every authenticated dashboard account. Uninstalling a harness preserves project data.
Remote Harness Clients
Use the model on this AI server from OpenCode, Pi or DeepSeek Harness running on another computer.
Run harness on this PCRemote AI / local tools
RECOMMENDED FOR LOCAL PROJECTSThe model runs on the AI server. File editing, terminal commands, Git, builds and tests run on the client computer against its local project folder.
Run harness on AI serverRemote AI / remote tools
SERVER WORKSPACEThe model, harness, files and tools all run on the AI server. Use the authenticated browser harnesses or SSH ai-code/ai-pi.
The exported connection profile contains server/model/harness information but never an API key. The Windows Remote Coding Setup stores its API key encrypted with Windows DPAPI. For access away from the LAN, connect Tailscale/WireGuard/VPN; do not expose raw model or harness ports with router port forwarding.
OpenCode Server
Shown when OpenCode is installed.
The server-hosted OpenCode web UI is protected by your AI Server Control login and uses the unified local model endpoint.
Shared Core Development Environment
Mainstream language runtimes, compilers, package managers and CLI verification tools shared by OpenCode, Pi and DeepSeek Harness. Heavy background/context features are optional below.
Checking coding environment…Live status from the shared Ubuntu coding toolchain.
Language Intelligence + Formatting
CLI lint/typecheck/test/build is the default diagnostic path. Language servers and automatic post-edit formatting are optional because they can add memory/CPU work.
Memory + Model Rules
Durable coding memory uses the shared SQLite/Git-aware service with bounded automatic retrieval. The legacy Markdown memory remains editable/compatible, while detailed language guidance stays in on-demand skills.
Shared Coding Memory
Long-term project memory for the shared OpenCode, Pi and DeepSeek workflows. Git and source stay authoritative; only retrieved context consumes model tokens.
ENABLED
Recommended default: enabled, 6K token memory budget, 12 retrieved items, 70% checkpoint target, FTS/BM25 + trigram + Git + repo map on, semantic embeddings off. The database and indexes consume no model context by themselves.
Memory Browser
Edit, pin, supersede or remove durable project knowledge. Source history remains separate.
| Memory | Type | Status | Importance | Updated | Actions |
|---|
Live Coding Performance Impact
Measured from this server so you can see what optional coding features are actually costing right now.
Loading real system measurements…
Optional Coding Performance Add-ons
Nothing here is required for normal coding. Language/toolchain packs are shared; OpenCode-specific toggles are only used by OpenCode.
Performance impact labels refer to runtime/context impact, not just disk space. Add-ons are shared server-wide and therefore only Administrators can install or toggle them.
Drive and folder review
Select any mounted SSD or other drive, or jump to a home folder. Includes hidden entries, small files and folder totals.
Changing this location also selects it for category cleanup and the largest-file scan below.
Review is read-only. Open folders to narrow the scan, or Manage to inspect and explicitly remove items in Files. Folder totals marked “on disk” measure allocated space; other sizes show apparent file data. Hard links and sparse files can change reclaimable space. Links are not followed and nested mounts are skipped. A large folder or an old file is not necessarily junk.
Choose a drive or home folder and review it.
| File or folder | Type | Size | Modified | Review guidance | Actions |
|---|
Cleanup preferences
Selecting a drive limits category cleanup to that filesystem, including its bind-mounted folders. Selecting a home limits cleanup to that home.
Scan known disposable files below, or review home folders for other clutter. Models, projects, conversations, credentials, ComfyUI inputs and model build caches are kept out of category cleanup. Nothing runs automatically.
Scan files to calculate amounts using these age limits and keep rules.
One full server path per line. Folder exclusions also keep everything inside them.
Review files
Scan to see files eligible for cleanup.
File sizes are estimates of reclaimable space. Changed, linked or open files are kept. A preview is valid for 30 minutes.
| Select | File | Category | Size | Modified |
|---|
Administrator login required.
Largest files
Scan any server folder and rank the files using the most space.
The scan is read-only and does not follow symbolic links. Whole-server scans skip virtual runtime folders. File sizes may not equal reclaimable space when a file has multiple hard links.
Run a scan to find large files.
| File | Size | Modified | Links | Actions |
|---|
Cleanup history
The last 20 runs, including any files kept and the reason.
Open the file manager to browse.
| Permissions | Owner |
|---|
No files in this folder.
Double-click to open. Use Ctrl or Shift to select several items, Ctrl+A to select all visible items, F2 to rename, or right-click for more actions. Drag files onto the list to upload. Text files up to 2 MiB can be edited here.
Disk
—
Server cleanup
Configure file categories and age limits, preview each file and choose what to delete.
Windows Model Library
Permanent model archive on a Windows hard drive with automatic fast-server caching.
On the Windows computer that owns the hard drive, install/update the Windows client and run AI Model Library Setup. It creates a restricted SMB share and registers it with this server. Model downloads can then write directly to the Windows drive. In Automatic mode, active models are copied to server cache when space allows; least-recently-used cache copies are evicted while preserving the configured free-space reserve.
Usage by Category
Partial Downloads
Probable Duplicate Large Files
Detected using identical size plus matching first/last 4 MiB. Nothing is deleted automatically.
Unused / Not Recently Loaded Models
Based on AI Server Control load history. “No recorded use yet” means tracking has not seen the model loaded since v7 was installed.
Configuration Backups
Configuration, accounts, workflows, custom nodes and version pins — not huge model files.
| Backup | Created | Size | Action |
|---|
Choose GPUs for ComfyUI
Tick one or more free GPUs. Each selected GPU starts a separate ComfyUI instance with its own queue and Open button.
Checking available GPUs…
Checking GPU instances…
Loaded models keep their GPUs. Each instance has its own workspace and queue; all instances share models, inputs and one output folder. Instances stay running until you stop them. Stop keeps instances with running or queued renders.
Advanced: one instance using several GPUs
For workflows with nodes that explicitly choose devices. This exposes multiple GPUs to one process; it does not combine VRAM.
Checking multi-GPU option…
Download model to AI server
Copy the download URL from ComfyUI’s missing-model notice and paste it here. The AI server downloads the file into its shared ComfyUI library.
Use a direct public model-file link. The filename comes from the URL unless you enter one. Existing files are kept. After the download finishes, refresh ComfyUI’s model list or reload its page. Missing custom nodes are installed through ComfyUI Manager.
Destination: AI server → ComfyUI models. No download is saved to this browser’s computer.
Custom Nodes
| Node | Status |
|---|
ComfyUI Control
MiniMax Music Production Toolkit
Checking…
Music production, mastering, artwork and audio enhancement nodes. Install on the AI server, refresh ComfyUI with Ctrl+F5, then open an example workflow.
Installation restarts ComfyUI if it is running. Finish queued workflows first. Models download separately on the server through the toolkit. The optional prompt writer needs llama-cpp-python and a compatible GGUF.
Toolkit and setup guide ↗ · Music workflow ↗ · Audio enhancement workflow ↗
Updates Git checkouts in custom_nodes, including disabled nodes. Stop every ComfyUI instance first. Local changes are preserved; restart instances to load updated code.
Curated Node BundleChecking…
What it does: installs the shared custom-node set used by the supplied image/video/audio workflows. Performance: no separate idle service; nodes only add runtime cost when a workflow invokes them. Measuring disk footprint…
SageAttentionChecking…
What it does: optional attention implementation for compatible ComfyUI/model workflows. Performance: no background process when idle; effect occurs only when a compatible workflow selects/uses it. GPU/VRAM impact depends on that workflow. Measuring installed footprint…
FlashAttention 2Checking…
What it does: optional optimized attention library used only by compatible models/nodes. Performance: no idle background process; CPU/GPU effect occurs during workflows that actually use it. Measuring installed footprint…
NunchakuChecking…
What it does: adds Nunchaku-compatible ComfyUI nodes/workflows for supported quantized diffusion models. Performance: no separate idle service; VRAM/speed changes occur only when a Nunchaku workflow/model is used. Measuring installed footprint…
Uninstalling these does not delete ComfyUI models, workflows, inputs or outputs. Removing the curated bundle leaves shared Python dependencies in place so other nodes are not broken.
Runtime Status
Blender + the official Blender Lab MCP.
Use Start & Open Blender above for the full editor in your browser. Pi and DeepSeek use the same scene. Start for models starts the saved runtime mode without opening the editor. Starting Blender requires an administrator account. Use Install / Repair first if needed. Save active work before stopping or restarting.
Model Connection
Pi or DeepSeek: start Blender above, open the harness from Launch connected apps, then ask it to “Use the Blender skill to create or edit a scene and save it under /srv/ai/blender-projects.” Pi uses the shared skill through ai-blender exec; DeepSeek uses the same fallback. Server OpenCode is configured with native MCP automatically. Any MCP-capable remote client can use the authenticated HTTPS endpoint.
Blender Settings
Save stores these settings. Graphics, GPU policy, workspace, startup file and online access apply at the next Blender start.
Loading settings…
Security: Blender MCP can execute arbitrary Python inside Blender. Treat model-generated Blender Python as code execution, save/version important .blend files, and keep remote MCP disabled unless you need it.
Internal bridge 127.0.0.1:9876 and local MCP 127.0.0.1:9191 are fixed to loopback for safety. Remote access uses the AI Server certificate and API-key authentication on 9445.
Memory breakdown & cache controls
Available RAM includes reclaimable memory. File cache helps reads run faster and Linux normally reclaims it when needed. These figures overlap; do not add them together. Application RAM is released when applications free it or close. Swap is disk-backed memory.
Open task manager to load Ubuntu processes.
| State |
|---|
Expand a group to inspect its processes. Group totals include every process in the group, even when searching. CPU is a share of the whole server (0–100%); ≥ means some processes are still sampling. RAM totals sum resident memory (RSS); shared pages can be counted more than once, so totals may exceed physical RAM. R = running, S = sleeping, D = waiting for I/O, T = stopped, Z = exited.
MCP services & connections
Shared services and MCP connections discovered in your coding harness settings.
Checking Godot…
Godot uses CPU graphics on a private browser desktop. Projects stay on Ubuntu. Editor/game windows keep running when MCP is turned off. Save projects before closing the desktop.
| MCP / owner | Control | State | Memory / CPU | Behavior |
|---|
Administrator login required.
Closing a browser tab does not close persistent server programs. Stop a service to release its process RAM. Save your work, then Close Godot desktop to release editor RAM separately. A harness connection change may require a reconnect or new session; it does not stop a remote application. RAM marked — is unknown or no local process was observed, not a zero-memory measurement.
Server
Network
GPU Detail
| GPU | VRAM | Load | Temperature | Power | Fan |
|---|
Server actions
Install tools, check health, or perform an explicit service or power action.
GPU Power Profile
All-card presets respect the individual wattage caps saved on Dashboard.
Thermal Protection
Scheduled Shutdown + Wake-on-LAN
Wake-on-LAN is enabled by the installer when supported. The Windows AI Server Control client can wake the server while the client is on the same LAN.
Select a log source and refresh.
Set up local speech
Install the speech and listening models once. Allow several GB of disk space and a few minutes for downloads. Your recordings and conversations are processed on this server.
Load your selected voice without starting the microphone. Start conversation also prepares it automatically.
Ready when you are
Tap once, speak naturally, and hear your model reply.
Microphone recording needs HTTPS (or localhost). You can always type below.
Off by default. When enabled, ask normally: public-information questions about weather, news, sports, prices and current facts can search automatically. Bing receives the current question and the server’s public IP; avoid private details in lookup questions. Recordings, saved voices and chat history are not sent. Say “don’t search” to keep a turn local. Lookups may retry once and add a few seconds.
Your conversation will appear here.
Your words, in your selected voice
Type exactly what you want to hear. This reads your text without asking the chat model.
Download WAVAI-generated speech · download to keep this recording
Memory & saved conversations
Continue a conversation
History is saved to your account on the server and is available on your phone and computer. New conversations can recall relevant earlier exchanges with the same voice.
Turning memory off stops earlier conversations and these notes from being used in new replies. Saved history stays until you delete it.
CREATE A VOICE
Let’s hear your voice
Use one speaker in a quiet room, at a comfortable speaking volume. Aim for 10–20 seconds without music or background voices.
Read this aloud for a useful sample
Hello! I’m glad we have a moment to talk. The morning was quiet, but by noon the streets were full of people. Shall we take a short walk, or make a fresh cup of coffee? Either way, there’s plenty to catch up on.
Read naturally in your usual accent. Keep the questions and pauses; do not rush to fit a timer.
Up to 30 seconds
Choose a recording, then prepare and listen to the exact sample before saving.
For Qwen, an accurate transcript enables transcript-guided cloning. Leave blank for speaker-only cloning. If you read the passage, enter only the words retained in the prepared sample.
A sample captures vocal sound. This description guides how the AI talks; it does not give it someone’s memories or personality.
TTS models · select or download speech models
Only supported text-to-speech and voice-cloning models are listed here. Chat uses the server’s active chat model, managed on the main Models page.
Models use their own local speech runtime, not the chat model loader. Qwen installs the complete native Safetensors repository, including its speech tokenizer. Existing chat-library downloads are preserved, not reused. Installation needs several GB of free disk space. Choose CPU or a separate free speech GPU in Conversation settings; CPU generation can be slow.
TTS Playground · test other engines
Test model and voice IDs outside the bundled list using a separately running local TTS server. It must support the OpenAI-compatible /v1/audio/speech API and return PCM WAV. This panel does not install engines, load raw model files, or change your saved voice’s conversation settings.
Administrator-only. Only your test text, model ID, voice ID and speed go to the connected engine—never recordings, saved voices or chat history. The engine manages its own voices, cloning, GPU and model loading. Stopping a test disconnects the request; the engine may still be generating.
Start a compatible TTS engine separately, then enter its URL and model/voice IDs. Model listing is optional; you can always enter an ID manually.
Download WAV
AI-generated speech. Download each sample to compare or keep it; server audio is temporary.
APPLICATIONS
Your programs
Launch an installed app, explore your tools, or review available updates.
Program updates Not checked yet
One click. Updates handled.
Checked daily. Nothing installs until you click Update.
GitHub release checks reuse verified results for up to 15 minutes. If GitHub limits requests, affected checks retry automatically after its cooldown; other update sources keep working.
Update checks the latest versions, waits for existing dashboard jobs, backs up configuration, installs and restarts affected apps automatically. Save unsaved coding and terminal work before clicking; connections can briefly drop. Models and account settings are kept. Held packages, custom files and failed checks are reported without being overwritten. Backups are not a full-system rollback. Once installation starts, do not shut down the server.
Recent program updates
Waiting for installed-program information.
Connect to your server to see its installed programs.
Your dashboard account opens ComfyUI, OpenCode and DeepSeek. With dashboard sign-in enabled, it also opens your own Open WebUI chats and memory. Cockpit and SSH use your Ubuntu login; shared files use your Samba credentials.
People & permissions
Manage who can use your AI server and who can administer it.
Server accounts —
Members can use AI tools. Administrators also manage people and the server.
| User | Role | Status | Last login | Manage |
|---|
AI Server accounts
Sign in to this dashboard, ComfyUI and the browser coding tools. New users choose their own password at first sign-in.
Ubuntu & shared files
Adding a dashboard user does not create an Ubuntu, SSH or Samba account. Those logins are managed separately.
Open WebUI
Dashboard sign-in gives each person separate chats and memory. Review, edit or delete remembered facts in Open WebUI Settings → Personalization. A server administrator enables dashboard sign-in with the Open WebUI setup launcher.
My Account
Appearance
Connection
The dashboard API is bound to the AI server itself and is only proxied through the same LAN/Tailscale web address.
HTTPS on this computer
Seeing “Not secure” on a new Windows computer? Set up trust for this server once for your Windows user.
Set up HTTPS on this computerExtract the download, run Set-Up-HTTPS.bat, and approve the Windows certificate prompt if shown. Then refresh this page. No additional server sign-in or Windows client installation is needed. Your browser cannot install certificate trust by itself.
Security
Remembered devices stay signed in until logout or access is revoked. Password resets sign out all devices. Clearing browser data, private browsing or browser cookie limits may require signing in again.
Tailscale / Network Administration
Network and Tailscale changes are administrator-only. Cockpit and SSH are also hidden from normal users because they bypass application permissions.
Tailscale Hardening
Device approval and Grants are enabled in the Tailscale admin console. AI Server Control deliberately does not store a tailnet administrator credential.
Power Management
Configure shutdown, idle power saving and how the Windows Power Bridge wakes the server while Ubuntu is off. BMC passwords stay on the bridge; its token is kept in a separate restricted secret file and is never returned to browsers.
Gigabyte BMC / Redfish
Automatic Idle Shutdown
Scheduled Startup
Startup runs on the always-on Windows Power Bridge using its local time, so it still runs while Ubuntu is off. Trusted Tailscale bridges need only their Tailscale URL; token-authenticated bridges also require their access token.
Scheduled Shutdown
Remote Start / Shutdown is served by the Tailscale-only Power Bridge because this dashboard is unavailable while Ubuntu is off. The BMC password remains on the bridge. Its access token is stored separately from dashboard settings with restricted permissions and is never returned by the API or written to audit details.
Recent Security Activity
| Time | User | Action | Address | Detail |
|---|
Live Data Sources
Production mode contains no simulated telemetry. If a metric cannot be read from the real server, it is shown as Unavailable.