A local-first desktop dashboard for building, running, and monitoring your Ollama AI models — all on your own machine.
Important
VELA is distributed as prebuilt release binaries only.
This repository hosts releases and documentation — the application source code is not publicly available.
To use VELA, download the latest build from the Releases page. There is nothing to clone, compile, or npm install.
VELA is a single, self-contained application that gives you a polished web dashboard for everything you do with local AI models. Point it at your local Ollama instance and you get a control center for your models: chat with them, build custom agents, monitor your GPU and system in real time, run a coding assistant with filesystem access, query your own knowledge base (RAG), and more — without sending your data anywhere.
It runs entirely on your computer. Your prompts, conversations, documents, and models never leave your machine.
- 🧠 Local-first & private — talks only to your local Ollama; nothing is uploaded.
- 📦 Zero install hassle — a single standalone executable. No Node.js, no dependencies to manage.
- 🖥️ Runs in your browser — start the app, open
http://localhost:3000. - 🎨 Modern, fast UI — dark/light themes, keyboard shortcuts, command palette.
Prerequisite: Ollama installed and running locally. Start it with
ollama serveand pull at least one model, e.g.ollama pull qwen3:8b.
- Download the build for your OS from the latest release:
- Windows:
vela-win-<version>.exe - Linux:
vela-linux-<version>
- Windows:
- Run it:
- Windows — double-click the
.exe. (See the SmartScreen note below — VELA is not code-signed yet.) - Linux — make it executable and launch:
chmod +x vela-linux-<version> ./vela-linux-<version>
- Windows — double-click the
- Open http://localhost:3000 in your browser.
That's it. VELA will detect your Ollama instance automatically. If port 3000 is busy, VELA picks the next free port and prints the URL in its console window.
By default VELA connects to http://localhost:11434. To use a different host, either set it in Settings → Connection (with a one-click Detect probe), or start with an environment variable:
OLLAMA_URL=http://192.168.1.100:11434 ./vela-linux-<version>- Agents — browse every local model in a table or card grid; sort by name, size, VRAM fit, or most-recently-used; star favorites; group models into named Collections.
- Build Agent — create custom agents with a guided form: base model, system prompt (with persona templates), full parameter tuning, and a live Modelfile preview. Hardware-aware recommendations suggest the right context size for your GPU.
- Pull & manage — pull models with a live progress bar, batch-queue multiple pulls, check the Ollama registry for model updates, and delete models from disk.
- Cookbook — a curated catalog of recommended models with VRAM-fit badges (
✓ Fits/⚡ Tight/✗ OOM) for your hardware.
- Chat — multiple session tabs, real-time token streaming with live tokens/sec and time-to-first-token, full-text search, and per-chat tool selection.
- Thinking blocks — native
<think>reasoning display for DeepSeek/R1 and Qwen3 models. - Vision & files — attach images for vision models and text files for context.
- Compare — send one prompt to two models side-by-side and diff their answers.
- Pipeline — chain models together so each step's output feeds the next, with forking and saved configurations.
- An Ollama-powered coding assistant with full filesystem access to a project folder: read, write, and edit files (with inline diffs), search, run shell commands, and more.
- Floating Terminal & Git windows — a built-in shell and a Git panel (status, log, diffs, commit, AI-generated commit messages) you can drag and resize.
- Works with any model — models without native tool-calling fall back transparently to prompt-based tools.
- Dashboard — at-a-glance model count, GPU VRAM/temp/utilization with sparklines, RAM, and active agents.
- System Monitor — live GPU, CPU (per-core), RAM, disk I/O, network I/O, model-storage usage, power draw, and a Live Agent Debugger that shows in-flight requests as they stream.
- Command Log — a timestamped record of every pull, build, and chat request.
- RAG / Knowledge Base — index text, files, URLs, or whole folders into a built-in local vector store (no external database). Semantic or hybrid BM25+vector retrieval with an optional AI reranker. Toggle it into chat for grounded answers.
- Documents — upload PDF, DOCX, XLSX, and TXT files; text is extracted and can feed RAG.
- Notes & Tasks — a built-in two-pane note editor and a task manager with priorities, due dates, and subtasks.
- Image Gallery — organize images with tags and notes for use with vision models.
- OpenAI-compatible API at
http://localhost:3000/v1— point Continue.dev, Cursor, Open WebUI, or any OpenAI SDK at your local models. - MCP tool server — expose VELA's model tools to Claude Code and other MCP clients.
- CLI — a bundled
velacommand for listing, pulling, chatting, and checking status from a terminal. - Fine-tuning — guided LoRA fine-tuning workflow (Unsloth → GGUF → Ollama) for supported setups.
- Keyboard shortcuts (
Ctrl+1–0), a command palette (Ctrl+K), dark/light themes, an interactive tutorial, and a built-in Wiki documenting every panel.
VELA checks GitHub Releases for new versions automatically (at startup and every few hours) and shows a "What changed" changelog right inside Settings → Updates — so you can read the release notes for each version, with anything newer than your installed build highlighted.
On supported Windows/Linux builds, VELA can download a new version in the background and apply it on the next restart. You can also always grab any build manually from the Releases page. Pre-release (-alpha/-beta) builds are included by default while VELA is in alpha — you can turn that off in Settings → Updates.
VELA is not code-signed yet, so the first time you run a new build, Windows SmartScreen may show a "Windows protected your PC" warning. This is expected for unsigned apps. To run it, click More info → Run anyway. (Code-signing is on the roadmap.)
| OS | Windows 10/11 (x64) or Linux (x64) |
| Ollama | Installed and running (default http://localhost:11434) |
| Node.js | Not required — the release binary bundles its own runtime |
| GPU (optional) | For live GPU stats: NVIDIA nvidia-smi, or AMD rocm-smi/amd-smi in your PATH |
Everything else (models, documents, notes, settings) is stored locally next to the app or in your browser — there is no database and no cloud account.
Ollama not detected
- Make sure Ollama is running (
ollama serve). - Check the URL in Settings → Connection and use ⌕ Detect to auto-probe common ports.
- If you set
OLLAMA_HOST=0.0.0.0with no port, Ollama defaults to port 80 — useOLLAMA_URL=http://0.0.0.0:11434explicitly.
The app window opens but the page won't load
- VELA prints its actual URL in its console; if port 3000 was taken it will be on a higher port (e.g. 3001). Open the URL shown there.
A login screen appears
- Optional password protection is enabled in Settings → Security. Enter your password, or disable auth if you're on a trusted local machine.
GPU stats are missing
- NVIDIA: ensure
nvidia-smiis in your PATH. AMD: install ROCm sorocm-smi/amd-smiis available.
Code Agent isn't creating files
- Pick a chat-capable model and make sure the working directory exists and is writable (click 📁 to validate). Tool-friendly models like
qwen3,llama3.1, ormistral-nemowork best.
VELA is local-first. It communicates only with:
- your local Ollama instance, and
- GitHub Releases — solely to check for app updates and show release notes (unauthenticated, no account, no tokens).
Your prompts, chats, documents, and models stay on your machine. VELA contains no telemetry or analytics.
VELA is closed-source. The application source code is not included in this repository or in the release downloads — only the prebuilt binaries and documentation are distributed here. Please obtain VELA exclusively from the official Releases page.
© ChinadaCam. All rights reserved. VELA is provided "as is", without warranty of any kind. Ollama and all third-party model names are trademarks of their respective owners.
⬇ Download the latest release · Built for people who run their AI locally.