Draft — not yet reviewed
A lawyer has not reviewed this text yet, and it may change. Placeholders in [brackets] will be filled in before launch.
Legal
Security
Updated 2026-11-08
How Mondoword handles your text: what leaves Figma, where it is stored, who can reach it, how long it is kept, and what AI agents may do.
Mondoword stores the text of your designs and sends it to AI providers to translate it. This page explains how that works and what protects it. The full list of services that handle data for us is on Where your data goes.
How your text flows
- Figma → plugin. The plugin runs inside Figma. It reads the text layers you link, select or translate, and the layer names and text of the frame they sit in. It never reads images of your designs, coordinates or fonts.
- Plugin → our servers. Every request goes over HTTPS and carries your sign-in token. Spell check is the one exception: it goes from the plugin straight to LanguageTool.
- Our servers → database. Keys, translations, glossaries, style guides and frame context are stored in our Postgres database.
- Our servers → AI provider. To translate, we send the text, key names, target languages, glossary terms, style guide and the frame's layer names and text to Anthropic's Claude API (the default) or Google's Gemini API (the fallback, and for frame summaries and glossary suggestions).
- AI agents. If you connect an agent through the MCP connector, it uses the same servers and gets only what you allow in Agent access.
Where data is stored
- Database: Neon Postgres on AWS in Frankfurt, Germany (eu-central-1). Point-in-time backups go back 7 days.
- Servers: Google Cloud Run. Production runs in Belgium (europe-west1).
- AI providers: Anthropic and Google process requests in the United States. [Owner: confirm]
- Integration tokens, such as your Crowdin token, are stored encrypted.
How long we keep it
- Your content stays until you delete it or the project.
- Server logs are kept for 30 days (the Google Cloud Logging default). We do not write AI prompts, AI responses or your texts to the logs in normal operation. Error logs can include short messages.
- AI usage records keep token counts per request, not text.
- AI providers process text to return a result and keep it as set by their API terms. Neither provider's API terms allow training their models on what we send. [Owner: confirm, including that the Gemini API runs on the paid tier]
Who can access it
- Sign-in with Google. We never see or store a password.
- Projects have members with roles. Today the person who creates a project is its owner and only member. Inviting teammates with a role per project comes with the web app.
- Sessions can be revoked. Beta access is checked every time the plugin opens. A revoked session or account stops working in the plugin and the MCP connector, at the latest when its sign-in token expires (within an hour).
- Our team: [Owner: describe who can reach production data, and how access is granted and logged.]
What AI agents may do (MCP)
- OAuth sign-in through Auth0 (EU), with the same Google account you use in the plugin. There are no API keys to copy or leak.
- Per-project capabilities. In the plugin's Agent access you choose, per project, whether an agent may view, edit or delete — or turn it off.
- Preview first. Every change an agent makes is shown as a preview first; nothing is saved until the agent confirms it in a second step.
- Two-step deletes. Deleting keys or translations and unlinking layers need a confirmation code from our servers, so the agent has to ask you first.
- Tokens are for us only. A connector token is accepted only if it was issued for our connector, belongs to an existing account, and was not revoked.
- Limits: 60 calls a minute (20 writes) per client, at most 50 items per call and 500 keys per translation job.
Sub-processors
The companies that process data for us, what each receives and where, are listed on Where your data goes.
Certifications
Mondoword has no SOC 2 or ISO 27001 certification yet.
Report a security issue
Use the contact form and choose Something isn't working. We reply within [Owner: response time].