Earlier posts in this series:
- A non-engineer CEO tried to connect Copilot Studio to LDX hub. Here's what happened.
- A non-engineer CEO connected Copilot Studio to LDX hub. It took 8 errors and 3 hours.
- I skipped Power Automate and connected LDX hub directly via MCP. It worked in 2 hours.
- I batch-processed 20 meeting minutes with Power Automate + LDX hub.
- We fed meeting minutes to AI without structure, and lost 70% of the data we needed for decisions
- Building a Technical Literature Dashboard with Power Automate, Power Apps and LDX hub StructFlow
I'm the CEO of a company in Tokyo that builds LDX hub, a language AI platform. I'm not an engineer. In May I wrote a series about connecting Copilot Studio and MCP to LDX hub. It took 8 errors and 3 hours the first time.
Since then, I packaged LDX hub as a plugin for AI agents. This post covers what's in it and what happened when I ran the same recording through three different agents.
Why a plugin
- Wiring each user's tool to our API by hand doesn't scale. My own first attempt proved that.
- I needed the agent to follow a process: show the cost and wait, ask a person to confirm names and terms, then run. A skill (a SKILL.md plus scripts) fits that better than a bare tool.
- RefineLoop reviews and revises a translation in repeated passes. A script handles that loop more reliably than MCP tool calls.
What's in it
Two editions. The GitHub edition is open to anyone. The Lab edition is for LDX hub Lab members and adds three things. Both are Apache-2.0. TermFit isn't in the GitHub edition yet, but you can use it if you contact us (https://ldxlab.io/contact).
| Skill | What it does | GitHub | Lab |
|---|---|---|---|
| LDX Voice | Transcribes recordings or Teams/Zoom transcripts, translates them, outputs SRT/VTT | Yes | Yes |
| LDX Voice dubbing | Synthesizes the translation as audio timed to the subtitles | No | Yes |
| RefineLoop | Reviews and revises a translated XLIFF until the output settles | Yes | Yes |
| StructFlow | Runs one instruction on every row or document, adds results as columns | Yes | Yes |
| StructFlow Loop | Checks documents or bilingual files against a style guide, MQM, or your own rules | Yes | Yes |
| TermFit | Applies a bilingual glossary to an XLIFF and fixes each mismatch | No | Yes |
| StructDraw | Pulls title blocks and parts lists out of drawing PDFs and DXF files | No | Yes |
| LDX Concierge - KŪ | Picks the right skill, estimates cost, tracks jobs | Yes | Yes |
How it works
You ask the agent in chat. The agent follows the skill's SKILL.md and runs a Python script on your machine. The script calls the LDX hub API with your key and returns the result files plus a comparison report.
- The scripts use only the Python standard library. Your API key stays in a local .env file.
- Before any paid step, KŪ shows an estimate and waits for your OK.
- At the end, a person reviews the comparison report and decides.
The test
One Japanese recording, 1 minute 3 seconds, plus a small glossary. I asked each agent for English subtitles and ran everything from the estimate to the final subtitles.
| Claude Cowork | Google Antigravity | OpenAI Codex CLI | |
|---|---|---|---|
| Edition | Lab | Lab | GitHub |
| Glossary terms matched | 11/11 | all | all |
| Max subtitle CPS | 17.8 | 17.8 | 17.8 |
| Dubbing | all blocks fit | done (re-synthesized at 1.074x) | not included |
| Waited for cost approval | yes | yes | no |
Models used: Google for transcription, Gemini 3.8 Flash for picking term candidates, Claude Haiku 5.5 for normalization, Claude Sonnet 5.5 for translation, Azure for dubbing.
What worked. The transcript had 9 misrecognized names and terms. All 9 were caught as glossary candidates and fixed after a person confirmed them. Review and revision settled on the second pass. Across the three agents, the translations differed only in wording.
What didn't. Codex started the paid transcription without waiting for my approval. Same instructions, different behavior. The fix is to make "wait for the reply" more explicit in the skill and to repeat the rule in AGENTS.md, which Codex reads. The same thing happened with our TermFit skill in Codex. After I rewrote that step, the retest stopped at the cost prompt as intended.
Also, about 6 seconds of one segment went missing in transcription, in all three environments. Re-transcribing just that segment brought it back.
Where it runs
Claude (Cowork and Claude Code), Antigravity, Codex CLI, and Cursor. There is also a Copilot Studio edition for Microsoft Power Platform with LDX Voice, RefineLoop and StructFlow.
It runs on your machine only. Your API key stays in a local .env file. That also means it won't run in browser chats like claude.ai or ChatGPT, because there is no safe way yet to hand the key to a cloud runtime.
Getting started
- Get an API key from the public gateway (https://gw.portal.ldxhub.io/api). You agree to the terms, AUP and DPA there.
- Get the GitHub edition from https://github.com/ki-dx/ldxhub-plugins, or the Lab edition from the download page, and add it to your agent.
- Put LDXHUB_API_KEY in the .env file in your working folder.
- Ask the agent: "What can LDX hub do?"
Next
LDX Voice in detail, then one skill a week: RefineLoop, TermFit, StructFlow, and StructFlow Loop.
A Japanese version of this post is on Zenn: https://zenn.dev/kozo_ki/articles/41029fe69f8d7d
LDX hub is a language AI platform built and run by Kawamura International in Tokyo. Plugin: https://github.com/ki-dx/ldxhub-plugins · Case studies: https://ldxlab.io
Top comments (0)