LLM Controller CE is a self-hosted, local-first browser app for operating local GGUF models: chat-first interaction, admin-controlled model discovery, loading, stopping, runtime settings, live logs, GPU/runtime visibility, analytics, benchmarks, installation, and user administration. Validated on Windows and Ubuntu when a compatible This is not just a launcher. LLM Controller CE is the Community Edition foundation of LLM Controller, built to make self-hosted local model operation feel like a real product. |
|
Running local models often means juggling folders, terminals, runtime flags, scattered utilities, and disconnected interfaces.
LLM Controller CE changes that.
It combines the operational side and the everyday usage side of local GGUF model operation into one unified web interface, including:
- Persistent Chat Workspace with sidebar search
- Reasoning-Aware Responses
- Managed Model Library
- File-Aware Conversations
- Controlled OpenAI-Compatible API Access
- Benchmarks
- Runtime Visibility
- GPU Monitor
- Installation & Settings
- Authenticated administration and user controls
- No cloud service required
Stream responses live, stop generation mid-stream, regenerate the latest answer, edit the latest prompt, and keep conversations organized through saved chat sessions with auto-generated titles, a searchable chat sidebar, and direct chat URLs that preserve the active chat across page refreshes.
When a model returns reasoning content, the interface can expose it with a dedicated show/hide workflow instead of burying it behind raw output.
Scan a configured model directory for GGUF files, maintain a registry of available models, recognize complete split model sets, load and stop models, save runtime defaults, mark favorites, enable or disable entries, and control which models are allowed in benchmarks. Administrators can also maintain compact usage profiles and notes, review each model's recorded Max TPS, configure an optional multimodal projector, and select an enabled model for chat-title generation.
Attach supported text files and PNG, JPEG, or WebP images through the file picker, drag-and-drop, or clipboard paste. Attachments remain associated with saved conversations and are validated against server-side count and size limits. Image understanding requires a compatible GGUF model, a compatible llama-server build, and a valid projector configured for the active model; text-only chat remains unchanged when no image is attached.
Administrators can generate a single CE API key and explicitly enable an allowlisted API surface on the existing application server. It exposes only GET /v1/models and POST /v1/chat/completions, uses Bearer authentication, and serves only the active model. Streaming, non-streaming, and supported inline multimodal requests use the same private loopback llama-server runtime; model lifecycle and administrative routes are not exposed.
Run administrator-controlled benchmarks across eligible models, edit the benchmark prompt set, review best-run summaries, move between summary and detailed saved outputs, and clearly distinguish current results from stale ones after prompt changes.
Watch live llama-server logs, runtime status, active process visibility, system RAM used/total, GPU telemetry where available, and per-model analytics without needing a separate dashboard.
The GPU Monitor supports NVIDIA and AMD telemetry paths where local tools such as nvidia-smi, rocm-smi, or rocminfo are installed and compatible with the host environment.
A built-in two-step installation flow initializes the application, prepares the MySQL database, creates the first administrator account, and saves runtime defaults before normal app access is opened.
LLM Controller CE is designed to feel like a real local AI control product, not a loose collection of scripts and utilities.
It brings together runtime control, conversations, reasoning-aware UI, observability, benchmarking, and system administration into one self-hosted experience that stays on your own hardware.
For people who care about local AI and controlling their own stack, this is the experience the software should deliver.
LLM Controller CE v1.2 is developed and validated with Python 3.11.7 and MySQL 8.0.22. It expects a self-hosted environment with:
- Python 3
- MySQL
- A working
llama-serverruntime compatible with the host OS and hardware - Local GGUF model files
- For image understanding, a compatible multimodal model and projector (
mmproj) file - A writable install folder
- Local NVIDIA or AMD GPU telemetry tools where GPU visibility is expected
LLM Controller CE uses a first-run web installer.
Basic flow:
- Install Python packages
- Prepare MySQL
- Place your
llama-serverruntime - Place your local models
- Start the application
- Complete the two-step installation flow
- Restart the app cleanly
- Log in and begin using the app
See INSTALL.md for Linux and Windows installation guides.
LLM Controller CE is built for authenticated use, including shared environments, not just a one-off single-user shell.
Current access behavior includes:
- Email and password login
- Remember-me support
- Role-aware interface behavior
- Administrator-only management controls
- Forced password change flow for accounts created with temporary credentials
LLM Controller CE is licensed under the GNU General Public License version 3.0 (GPLv3).
See LICENSE for full terms.
LLM Controller CE is developed by Tensioncore Administration Services.
LLM Controller CE is the start of a broader product direction:
run local AI cleanly, monitor it properly, evaluate it honestly, and keep control of your own infrastructure.
That's what this project is about.