trina-app/trina-ax

What it does

Trina AX puts Claude Code and Codex agents on a company server instead of scattering them across individual laptops. Company files live in shared/, groups/, projects/ and personal folders; skills are plain SKILL.md files; and everyone reaches the agents through a browser chat, desktop or phone client. The intended audience is a team that wants one shared context for AI work without asking every employee to install and configure a coding agent.

The core design is permission parity between the human and the agent. Each file access is checked as the person asking, so a folder the user cannot open is also invisible to the AI across files, answers, search, previews, version history and trash. Agents run with their own shell and file tools disabled and reach company data only through Trina's checked file tools; engine tokens are replaced with per-turn stand-ins. On Linux and role-account Mac installs, OS-level isolation places the engine in a dedicated account, while the office Mac and local demo do not get that isolation (security and permissions).

Text is extracted from PDF, Office and HWP/HWPX files. Admins cannot browse private conversations, though server operators can read stored data. Several starter skills (weekly report, sales follow-up, brand brief, expense review, onboarding, meeting summary) illustrate the SKILL.md convention (examples).

Quick start

A 10-minute local demo needs Git, Bun 1.3.12 and Go 1.27+. It uses fictional data and runs with AI disabled, so no Claude or OpenAI account is required:

git clone https://github.com/trina-app/trina-ax.git && cd trina-ax
bun install --frozen-lockfile
bun run demo        # → http://127.0.0.1:19760

Sign in as alice with password teamax-demo-1, then sign in as bob in a private window to see a different group folder. The demo is explicitly not a production install; the full walkthrough is in the quickstart.

For production, choose Trina hosting, Linux self-hosting, or an office Mac install. The Linux guide covers Debian 12/13 or Ubuntu 24.04, systemd with cgroup v2, amd64/arm64, 4 GB RAM recommended and 20 GB disk minimum.

Usage constraints

• Self-hosting is not offline. Selected context is sent to the configured AI provider, and each Codex web search sends the conversation so far to OpenAI's search service. The server also polls updates.trina.app for its public address every 30 minutes unless MYCEL_EXIT_ADDRESS_CHECK=off is set.
• Linux servers run no OCR: scanned PDFs and photos are not text-searchable there. HWP/HWPX text extraction works, but visual previews require an optional plugin needing macOS or Linux with glibc 2.39+, which rules out Debian 12.
• Video and audio reading is Linux-only, and speech transcription requires an admin to enable it and goes through the team's configured speech provider.
• Removing folder access does not erase earlier AI answers, and app admins cannot read chats while server operators can read stored files and chats.
• Trina provides the workspace, not the AI subscription; Claude and OpenAI accounts must be authorized for the team's use, and that software connection setting does not grant permission under either provider's terms.
• AGPL-3.0-only for the application, MIT for the cloud protocol. Published installers can lag behind source changes.

HotRepo's take

The project appeared on HotRepo's new-project list for TypeScript on 2026-11-02, with 106 stars and 4 new stars in a roughly two-hour window. The star curve over that afternoon moved from 102 to 106, a small sample that supports only the day's activity, not a trend. What is notable from the material itself is the concrete permission model — per-call checks as the requesting person, with agent shell tools turned off — presented alongside a runnable AI-disabled demo, which gives readers a way to inspect the workspace without settling an AI account first.