_{English · Español}

_{Website · Documentation · Changelog · License}

A private, local-first assistant for your files, research, and code.

TrinaxAI runs on your computer and keeps inference and indexed data on the configured host unless you explicitly choose a remote service. It combines local chat, cited retrieval-augmented generation (RAG), optional web research, a workspace-confined coding agent, a terminal interface, and an installable progressive web app (PWA).

Install with npm

Use the npm package on Linux, macOS, or Windows. You need Node.js 22 or newer with npm available in your terminal.

npm install -g trinaxai@latest
trinaxai setup

The launcher downloads the matching official release, verifies its SHA-256 checksum, and starts the platform setup. The setup flow prepares the backend, PWA, Ollama, and recommended models.

Setup opens with a welcome and asks for your language (English or Spanish). It then checks your disk space and connection, asks every optional question up front (model downloads, start now, autostart, weekly update check, access from other devices), and shows an installation plan to confirm before it changes anything. Run trinaxai setup again on an existing installation to repair, update, or uninstall it.

The npm launcher is the only supported end-user installation path. The platform scripts in this repository implement the release behind the launcher; you do not need to download or run them directly.

Start using TrinaxAI

Run the health check first:

trinaxai doctor
trinaxai status

Then open https://localhost:3334. Your browser may ask you to trust the local certificate on the first visit.

Ask a question about indexed files:

trinaxai ask "Summarize my indexed project" --engine rag

If you skipped model downloads during setup, run trinaxai setup again when you are ready. Use trinaxai --help to see every available command.

Setup options

Need	Command
Defer model downloads	trinaxai setup --no-models
Prepare without starting services	trinaxai setup --no-start
Let other devices on your network open TrinaxAI (default: this computer only)	trinaxai setup --lan-access
Choose a hardware profile	trinaxai setup --profile 16gb
Preview changes	trinaxai setup --dry-run
Skip the language question	trinaxai setup --language es
Automate setup	trinaxai setup --yes (or --non-interactive)

The installer detects CPU, memory, GPU, and VRAM. The default profile is safe for the detected machine. See configuration when you need to tune models, resource limits, or providers.

Daily commands

Use the CLI from any directory after setup:

Goal	Command
Open the interactive assistant	trinaxai chat
Ask one question	trinaxai ask "..." --engine rag
Index a folder	trinaxai index ./documents
Run deep research (add --web to ground it on web search)	trinaxai research "..." --depth 2
Use the coding agent	trinaxai agent --workspace .
Check services	trinaxai status
See or change who can reach TrinaxAI	trinaxai network / trinaxai network off
Start or stop services	trinaxai start / trinaxai stop
Diagnose a problem	trinaxai doctor --strict
Update the installation	trinaxai update
Remove TrinaxAI	trinaxai uninstall

What TrinaxAI includes

• Local Ollama chat with task-aware routing for chat, code, reasoning, and math
• Hybrid RAG over code and documents with citations, collections, and incremental indexing
• Optional web search, deep research, memory, voice, and vision
• A coding agent with approved workspaces and action approvals. Its terminal tool runs in a networkless bubblewrap sandbox on Linux and stays disabled on macOS and Windows; the file tools work everywhere (see the security guide)
• An HTTPS PWA for desktop and mobile with scoped device pairing
• A terminal interface for chat, indexing, research, diagnostics, and exports

Models and hardware

Setup chooses a model profile from available memory and GPU resources:

Profile	Chat and code	Fast responses	Embeddings
8gb	qwen3.5:2b	qwen3.5:2b	qwen3-embedding:0.6b
16gb	qwen3.5:4b	qwen3.5:2b	qwen3-embedding:0.6b
32gb	qwen3.5:9b	qwen3.5:4b	qwen3-embedding:4b
64gb	qwen3.5:35b / qwen3-coder:30b	qwen3.5:4b	qwen3-embedding:4b

The 8gb profile works with CPU-only hardware. See the configuration reference for model and resource settings.

Known limits

• Questions about code answered by small local models: the cited sources are sometimes wrong or incomplete, so open the cited file before relying on an answer.
• trinaxai research is slow on a CPU-only computer.
• Not measured: the full flow (install, chat, RAG, agent, web app) on Windows, on a GPU, and with large models. macOS and Windows are covered by the automated CI checks, not by end-to-end runs.

Privacy and security

TrinaxAI binds local services to loopback by default. Ollama is not exposed as a generic proxy. LAN browsers must pair with a one-time code and receive only the capabilities you grant, device by device. You can let a specific device index, use the Agent, or modify data; power, model management, device pairing, factory reset and Agent auto-approve (YOLO) always remain host-only.

Keep ports 3333 and 11434 private, protect storage/.proxy_secret, and use a VPN instead of exposing the host to the public Internet. Read the security guide and LAN pairing guide before enabling another device.

Documentation

You want to…	Read
Install, update, or remove TrinaxAI	Documentation hub
Configure models, RAG, networking, or limits	Configuration
Look up an environment variable	Environment variables
Use every CLI command	CLI reference
Understand the architecture	Architecture
Integrate the HTTP API	API reference
Recover from an error	Troubleshooting
Pair another device	LAN pairing
Develop or contribute	Developer guide

The PWA also exposes these guides under Settings → Documentation.

Development

This section is for contributors working from a source checkout. End users should use the npm installation above.

git clone https://github.com/TrinaxCode/TrinaxAI.git
cd TrinaxAI
python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.lock
(cd chat-pwa && npm ci && npm run dev)

Run make check before submitting a change. See the contributing guide for the full workflow.

Screenshots

Flow	English	Español
Chat with citations	Open	Abrir
Indexing job	Open	Abrir
Pairing	Open	Abrir
Agent approval	Open	Abrir

License

TrinaxAI is licensed under AGPL-3.0-or-later. See LICENSE and the trademark guide.

Built by TrinaxCode · trinaxai.app