Skip to main content
Glama

The AI your whole team shares. Trina AX puts your team's files, skills and tools on one server. Everyone's AI starts from them,and every file it opens is checked as the person who asked.

       Linux: .deb and AppImage (x64, arm64) on the download page

Website · Docs · Pricing · Discord

English · 한국어 · 日本語 · Español · 简体中文

https://github.com/user-attachments/assets/a918709d-019d-4743-8e31-f05dd1b4e2ca

Set up once. Your whole team gets it.

One AI account for the whole team. Connect one Claude or OpenAI API key, or a plan whose terms allow it, and everyone asks with it, with token limits per person, group or team.

One person sets things up on the team server. Everyone's AI has it from their first question, in the browser or the apps. Nobody else installs Claude Code or copies skills; at most they sign in to their own accounts.

Set up once

Everyone gets

AI account. The owner or an admin connects a Claude or OpenAI API key, or a plan whose terms allow it.

Answers for the whole team, with token limits per person, group or team.

Google. Share Drive folders with the team, or connect one team account (a shared mailbox or calendar) for Gmail, Calendar, Docs and Sheets.

The people you choose use them in their chats. Anyone can still connect their own Google account.

Apps. Higgsfield, Notion, Linear, Figma and more from the app list, or any MCP server by its address, such as Bright Data.

The app's tools in everyone's chats: one team sign-in, or each person signs in as themselves.

Your company API. An address, a key and a few notes.

The AI calls your own system, only under that address.

Slack. The owner or an admin adds it once.

Ask Trina in a DM or any channel it's invited to, with your own access.

Skills. Write a SKILL.md, or bring the skills in your .claude or .codex folder.

Everyone's AI follows the same steps. Type / to use one.

Team files. One drive: company-wide, groups, projects, team skills and personal folders.

Everyone's AI starts from the team's files, and opens only what that person may open.

Related MCP server: AI News Aggregator

What everyone does with it

Ask beside any room, send in one click. A side chat only you see; one click sends Trina's answer into the room.

Each person's AI sees only what they may open. Mina can't open Finance, so neither can her AI. How permissions work

Pick up a teammate's session in a team project. Sessions in a team project are shared with its members: open a teammate's and keep going, with your own access. Any other session you can share with people or groups, like a file in Google Drive.

Files found by what they are. Every night the team's AI reads new and changed files in shared, group and project folders and keeps one line per file saying what it is, so a scan named temp.pdf or a photo named IMG_2031.jpg is found by its content. It uses the team's AI account and never reads personal folders. In a test with 2,238 made-up files, Claude Code found the right file for 181 of 184 questions with these lines, and for 172 without them.

  • Every change is kept. By a person or their AI; bring back any version.

  • In-app browser. In the desktop app, Trina opens websites in a pane beside your chat, reads them, clicks and fills in forms while you watch; Stop ends it. It asks before a new site unless you allow all websites.

  • Apps post into rooms. Sentry, GitHub, Stripe, Jira and HubSpot updates arrive as cards.

  • Find work files. The desktop app finds recent work files on your computer and adds the ones you pick.

  • Documents. Reads Hangul (HWP/HWPX) text; previews PDF and PowerPoint.

  • Every device. Mac, Windows, Linux, iPhone, Android and any browser, in six languages.

Get started

Self-host with Docker (Linux, or a Mac with OrbStack)

git clone https://github.com/trina-app/trina-ax.git && cd trina-ax/docker
cp env.example .env && docker compose up -d --build
docker compose exec trina-ax trina-ax-setup-link   # open the link and create the owner

Newer and less tested than the Linux install; an Ubuntu 24.04 server needs one host setting first. Docker guide

Try it on your computer in 10 minutes (fictional team, AI off)

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 · alice / teamax-demo-1

Needs Bun 1.3.12 and Go 1.27+. Not a production install. Walkthrough

How it works

One server runs Claude Code and Codex as the engine. Every file request, from a person or their AI, goes through one permission check on the server. On a Linux server the engine runs as its own account inside a bubblewrap, Landlock and seccomp sandbox, with its shell and built-in file tools turned off: it reaches files only through Trina's tools. Security and permissions

Trina provides the workspace, not the AI: connect Claude or OpenAI accounts and plans authorized for your use. AI usage is billed by the account you connect.

What we don't promise

  • Self-hosting is not offline: the context a request needs goes to the AI provider you connect, and configured integrations receive their requests. Each Codex web search sends the conversation so far to OpenAI's search service. With OpenRouter for speech, allowed by an admin, the sound of video and audio files goes to OpenRouter. The server asks updates.trina.app for its own public address every 30 minutes (MYCEL_EXIT_ADDRESS_CHECK=off turns that off). Details

  • Server operators can read stored files and chats, including Trina operators for hosting. App admins don't get a browser for everyone's private conversations.

  • Removing folder access does not erase earlier answers. The wider-sharing check is not universal data-loss prevention.

  • No security certification, perfect answers, guaranteed savings or automatic legal, financial or HR decisions. Review outputs and deployment boundaries.

  • Linux servers run no OCR: the AI sees photos and scanned pages as images, but they are not text-searchable. Office Macs cannot read video yet.

  • HWP/HWPX text works for every team; previews and making Hangul documents need the optional HWP Preview plugin (macOS, or Linux with glibc 2.39+; not Debian 12).

  • Connections: a person's own eligible connection comes first, then one assigned to the person, a group, then the whole team. On fresh Trina teams, one company connection can serve the whole team by default; existing saved choices persist. This software setting does not grant permission under Anthropic's or OpenAI's terms. Connection details

  • Linux self-hosting needs Debian 12/13 or Ubuntu 24.04, systemd with cgroup v2, amd64 or arm64, 4 GB RAM and 20 GB disk. Your IT team installs, updates and restores backups.

Community

If Trina AX helps your team, a ⭐ on GitHub helps other teams find it.

Discord · X · Releases · Changes · Support · Contributing · Security · Development · Configuration · llms.txt

The app speaks English, Korean, Indonesian, Japanese, Spanish and Simplified Chinese. Fictional company and people in every picture.

License

Free to use, including commercially, under AGPL-3.0-only. If you offer a modified version over a network, offer its source to those users. The cloud protocol is MIT (license). Notices · Third-party notices

Claude Code and Codex have their own terms; this project is not endorsed by Anthropic or OpenAI. Claude is a trademark of Anthropic.

Available Tools

8 tools
check_cacheA

Check the current state of the news cache — Supabase connection status, whether local/Supabase cache exists, how old it is, and how many items per source.

Inputs: no named parameters.

get_new_sinceA

Queries Supabase for all cache rows created after a given ISO timestamp. Returns deduplicated news items sorted newest first. Requires Supabase connection.

Inputs: since (required).

get_paper_briefA

Fetches an ArXiv paper's title, authors, abstract, and submission date. Searches Papers With Code and GitHub for associated code repositories.

Inputs: url (required).

get_repo_quickstartA

Fetches a GitHub repo's metadata (stars, description, language, topics) and extracts install commands and a quickstart code block from its README.

Inputs: url (required).

get_topic_suggestionsA

Suggests blog topic ideas based on trending news, tailored per project (potenlab, treesoop, hanguljobs) and filtered against already-used topics.

Inputs: project (required), slots, used_topics.

get_top_picksB

Returns top N most relevant items for AI engineers, scored by source reputation (HN/Reddit > ArXiv > others) plus item score. Each item includes a one-liner 'why it matters' and optional try_url.

Inputs: n, category.

search_todayA

Search today's cached news by keyword query. Matches against title and summary, scores by word match count × item score. Returns top results.

Inputs: query (required), limit.

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    TDQS

    A3.9/5.0

    Scored across 8 tools

    Disambiguation4/5

    Tools like get_trending_news, get_top_picks, search_today, and get_new_since all return news items but with distinct mechanisms—aggregation, ranking, keyword search, and time-based filtering. Some initial confusion possible between get_trending_news and get_top_picks, but descriptions clarify their unique purposes.

    Naming Consistency4/5

    Most tools follow a get_ prefix with descriptive nouns (get_repo_quickstart, get_trending_news, get_top_picks, get_paper_brief, get_new_since, get_topic_suggestions), but search_today and check_cache deviate from this pattern. Still, the naming is largely predictable and readable.

    Tool Count5/5

    With 8 tools covering news aggregation, search, ranking, paper/research lookups, repository quickstarts, topic suggestions, and cache status, the tool count is well-suited to the server's purpose. No tool feels redundant or unnecessary.

    Completeness4/5

    The tool surface covers the main workflows: retrieving trending news, searching it, getting top picks, fetching paper briefs and repo quickstarts, suggesting blog topics, and checking cache health. Minor omissions like fetching a specific news item by ID or source-specific filtering are workaround-able through search and time-based retrieval.

    Maintenance

    ActivityNo data
    ResponsivenessLimited data

    Related MCP Connectors

    Related MCP Servers

    • F
      license
      Not graded
      quality
      D
      maintenance
      Local MCP server that ingests and normalizes recent AI, tech, and research news from multiple sources, exposing tools for media agents to query and retrieve article content on demand.
      -
    • A
      license
      Not graded
      quality
      D
      maintenance
      MCP server that aggregates AI news, academic papers from ArXiv, and trending GitHub repositories from multiple sources, enabling users to fetch, search, and filter recent AI content via natural language.
      1
      MIT
    • A
      license
      A
      quality
      D
      maintenance
      Provides AI agents with global tech news from 500+ sources across regions, with translated titles, scoring, and clustering tools for real-time intelligence.
      8
      MIT
    • A
      license
      Not graded
      quality
      B
      maintenance
      MCP server that gives AI agents a ranked, evidence-backed board of what's trending in AI — repos, models, papers, news — and a daily brief, using fused signals from GitHub, Hacker News, npm, and more.
      132 PyPI
      4
      Apache 2.0