Community
Self-hosted
$0forever
₩0forever
- The full product
- Every security feature
- Unlimited people, folders and connections
- Community support on GitHub
Open-source release coming soon
Give your team’s AI agents your company files.
They only see what each person is allowed to see, never touch originals, and everything is logged. Open source, on your own Mac mini, over the folders you already have.
tdrive mcp add claude Demo transcript
- # kim’s Claude Code — kim is not in 경영 (management)
- kim $ ls /
- 개인_kim/ 공용/ 프로젝트/
- kim $ grep -ri revenue /
- (no matches — for kim, 경영 does not exist)
- # boss’s Claude Code — boss is in 경영
- boss $ grep -ri revenue /
- /경영/이사회_메모.pdf: Revenue target: 12.4B KRW
- boss $ echo "Q3 summary …" > /경영/q3.md
- Originals are read-only. Write to /drafts/ and a person will move it.
- boss $ echo "Q3 summary …" > /drafts/q3.md
- ✓ Saved to /drafts/q3.md — contains content from 경영
Works with the AI tools your team already uses
- Claude Codetoday
- Codextoday
- Cursortoday
- claude.aiv0.2 · rolling out
- ChatGPTv0.2 · rolling out
The problem
Today an agent sees all of your files, or none of them.
All or nothing
A synced folder gives the agent everything that laptop can open. There is no way to give the intern’s agent just one project folder.
Originals at risk
An agent with write access edits the real quote sheet. Nobody can tell which agent changed what.
No trail
The Claude Code on someone’s laptop leaves nothing the team can look at later.
HWP and PDF are a black box
grep does not reach inside the PDFs, Office files and HWP documents where the actual work lives.
Trina Drive sits between every agent and your folders: one permission filter, one draft box, one audit log.
What you get
Agents work the way they already do. People stay in control.
-
One permission filter for every path
Shell, file tools, search and the web all ask the same function. An agent gets the person’s rules narrowed by its token’s scope. What it may not see does not exist: same output, same error, same exit code.
-
Read-only originals, drafts with provenance
Agents write only to /drafts/. Every draft carries the folders it drew from, and only tokens that can still read all of them see it. Moving a draft somewhere wider goes through a person.
-
An isolated shell agents already know
ls, grep -r, cat and pipes over a virtual tree. Each command runs in a sandboxed worker with time and memory limits: no network, no real processes. grep is real ripgrep.
-
Full audit, with signed anchors
Who, which connection, which command, which files — in a hash-chained log. The chain head is signed and sent out daily, so rewritten history is caught. If the log cannot be written, the call fails.
-
On your NAS folders and your Mac mini
Register the SMB shares or local folders you already have. Nothing moves. People keep using Finder, Explorer, Office and Hangul. Runs on a Mac mini in the office, or Linux with Docker.
-
Full-text grep in HWP, PDF and Office
grep -r reaches inside PDF, DOCX, XLSX, PPTX, HWP and HWPX. Documents are parsed in isolated workers and the text is cached by content hash.
How it works
Three steps. No migration.
- 01
Install on a Mac mini, register folders
Run the installer, pick folders on your NAS, and answer two questions per folder: who opens it on the NAS, and whose AI may read it.
- 02
Each person connects their AI
One command makes a personal token and adds it to Claude Code, Codex or Cursor. 7 days by default, narrowed to the folders you pick, kept in the Keychain — never in a repo.
tdrive mcp add claude --server https://drive.office.lan --login kim - 03
Agents work. You see everything.
Agents read only what their person may read and leave drafts with sources. Every call lands in the audit log. Revoke a connection and the next request fails.
Your people and their AI
- Claude Code
- Codex
- Cursor
- claude.ai · ChatGPT (v0.2)
Personal token or OAuth, per person
Trina Drive on your Mac mini
- Permission filter (one function)
- Isolated shell · read · drafts
- Audit log + signed anchor
Storage, decisions and logs stay here
Your folders, as they are
- NAS shares — read-only
- Local folders — read-only
- /drafts/ — on the server
People keep using Finder, Office, Hangul
Security
Honest about what we protect, and what we can’t.
Storage, permission decisions and the logs are on your server. What an agent reads goes to the AI company that made that agent (Anthropic, OpenAI and others).
Mark folders where that must not happen as “No AI”. Personal folders cannot be read secretly, not even by admins (unregistering one or changing its owner notifies the owner), but whoever runs the server — for managed plans, our operators — can see the disk.
What protects you
- Agent access = the person’s rules ∩ the token’s scope, decided in one place.
- Hidden is the same as missing: same output, error and exit code.
- Nobody writes originals through Trina Drive — not even admins.
- Tokens are hashed, short-lived, kept in the OS Keychain and revocable at once; a new client triggers an alert.
- Shell commands and document parsing run in sandboxed workers.
- Bulk-read alarms; per-token rate limits; fail-closed audit.
What we can’t stop
- Once an agent reads a file, its AI company receives that content. Keep such folders “No AI”.
- Copies made through an agent’s memory or local files, or drafts a person rewrites heavily.
- Access that does not go through Trina Drive, such as a laptop mounting the NAS directly.
- NAS permissions are not read automatically: you tell Trina Drive who opens each folder on the NAS.
Pricing
Self-hosted free forever. Pay us to run it.
Every security feature is free and self-hostable. We charge for running Trina Drive for you, and for setup. Managed plans are priced on AI-connected people — everyone who only opens files is free.
14-day free trial of Team after a setup call. Annual billing = 2 months free. All plans are billed in US dollars (USD).
Team
Managed
$99/ month
or $990 / year
$990/ year
2 months free vs. monthly
₩129,000/ month
or ₩1,290,000 / year
₩1,290,000/ year
2 months free vs. monthly
- We install, monitor, update, back up and secure it on your office Mac mini (remote)
- Up to 10 AI-connected people
- Unlimited file-only people
- Extra AI-connected person: $9 / monthExtra AI-connected person: $90 / yearExtra AI-connected person: ₩12,000 / monthExtra AI-connected person: ₩120,000 / year
- Email support, next business day
Business
Managed
$349/ month
or $3,490 / year
$3,490/ year
2 months free vs. monthly
₩449,000/ month
or ₩4,490,000 / year
₩4,490,000/ year
2 months free vs. monthly
- Everything in Team
- Up to 40 AI-connected people (+$7 / month each beyond)Up to 40 AI-connected people (+$70 / year each beyond)Up to 40 AI-connected people (+₩9,000 / month each beyond)Up to 40 AI-connected people (+₩90,000 / year each beyond)
- 4-business-hour response
- Quarterly permission & audit review call
- Setup included on annual
Setup & migration
One-time
from $1,490
from ₩1,900,000
On-site in Korea available
- Folder & permission design
- NAS / Mac mini setup
- Client connection training
FAQ
Questions, answered plainly
Do my files leave the office?
Your files stay on your NAS and Mac mini, and storage, permission decisions and logs are on your server. But what an agent reads is sent to the AI company behind that agent — that is how AI tools work. Folders that must never reach an AI company can be marked “No AI”: agents cannot see them at all.
Which AI tools work with it?
Claude Code, Codex and Cursor today, over MCP with a personal token: tdrive mcp add claude (or codex, cursor). The token is written to your user settings and the OS Keychain, never into a project’s .mcp.json.
What about claude.ai and ChatGPT?
Connectors for both, over OAuth, are built and rolling out in v0.2. They need a public HTTPS address (a Cloudflare tunnel that exposes only the MCP, sign-in and consent paths) and a second factor for every admin. On the AI side: in claude.ai Team and Enterprise the owner adds the connector; in ChatGPT, MCP with writes needs a Business plan and developer mode.
Is it really open source?
The open-source release is coming soon, and the license will be announced before it. Our commitment: every security feature stays free and self-hostable, with no limits on people, files or connections.
Do you need my NAS?
No. Trina Drive runs on a Mac mini in your office and reads the folders you register — SMB shares on your NAS or folders on the Mac itself. Nothing is moved or copied to us. On managed plans we operate that Mac mini remotely; the NAS stays yours.
Can an agent change or delete our files?
No. Through Trina Drive nobody writes originals, not even admins. Agents write drafts to /drafts/, and a person decides where they go. Deleting originals is not possible through it.
What if a document tricks the agent (prompt injection)?
We cannot stop an agent from sending what it read to its AI company. We limit what it can read in the first place (per person, per token, “No AI” folders), keep originals read-only, label drafts with their sources, alert on bulk reads and log every call.
Someone left, or a laptop was lost. Now what?
Revoke their token on the web; the very next request fails. Tokens expire (7 days by default) and the owner is alerted when a new client starts using one.
Who counts as an AI-connected person?
Anyone who connects at least one AI tool to Trina Drive. People who only open files on the NAS or in the web app are free and unlimited.
Do people have to change how they work?
No. People keep opening files in Finder, Explorer, Office and Hangul. Only their AI tools go through Trina Drive.
Let your agents into the shared drive — only as far as each person goes.
Free to self-host. Or let us install and run it on your office Mac mini.