Security and data handling
Blog Agent writes to your repository and deploys your site, and it uses a large language model to do it. This page states what that model sees, what the agent can touch, and how to set it up so it can touch as little as possible.
At a glance
- Nothing is published without a human pressing Publish. There is no unattended mode.
- Drafting uses Anthropic’s Claude models. Your prompt, the repository files the agent reads, and the public sources it researches are processed by Anthropic.
- Articles are plain files committed to your own Git repository, so every change is in your version history.
- The review dashboard listens on localhost only and is not exposed to the network.
What the model sees
The agent runs Anthropic’s Claude through the Claude Code command-line tool. For each run it works in a copy of your website repository and reads what it needs there: existing posts, to avoid repeating topics, and the files it has to update, such as the sitemap. It also fetches public news sources for your industry.
It does not need, and should not be given, access to customer data, mailboxes or internal systems. The model only sees what is in the repository and on the public web.
Credentials it holds
| Credential | Used for | Recommendation |
|---|---|---|
| Git push to your website repository | Committing approved articles | A token or deploy key scoped to that one repository |
| Deploy rights in your cloud project | Building and deploying the site after approval | A service account that can build and deploy that one service, nothing else |
| Model access (Anthropic) | Research and drafting | A dedicated account or API key for the agent |
The agent runs commands on the machine it is installed on, with the permissions of the account it runs under. Give it a dedicated machine or user account and the scoped credentials above, and it cannot reach anything else.
Controls that stop bad output
- Human approval. Every article waits for review, and the agent will not start a second one while the first is pending.
- Quality gates. Word count in both languages, service links, meta length, keyword presence, matching metadata, sitemap coverage and a clean build. Any failure stops the deploy.
- Post-deploy checks. It fetches the live pages, confirms the article, language tags and sitemap, and reverts to the previous version if something is wrong.
- Words you approved stay the words you approved. Approved text is copied to your site unchanged.
On our own site, the article HTML is also sanitised when it is served, so markup that should not be in a post cannot reach a visitor. We recommend the same on yours.
Who builds it
Blog Agent is built by IDE Solutions UG (haftungsbeschränkt), Krombach, Germany (see the Impressum). In the fully managed option we run it for you under a contract that includes a data processing agreement under Art. 28 GDPR. See deployment options and pricing.
Security review questions
Does our content leave our infrastructure?
Partly, and it should be said plainly: drafting runs on Anthropic’s Claude models, so the prompt, the repository files the agent reads and the draft text are processed by Anthropic. The finished articles, the repository and the credentials stay with you.
Can the agent publish on its own?
No. Every article stops at human review. Publishing only happens after someone presses Publish in the dashboard.
What is the worst a compromised run could do?
Whatever the account and credentials it runs with allow. That is why it should run under a dedicated account with a repository-scoped token and a deploy-only service account: then the limit is one repository and one service.
Go through it with your team
We walk through the credentials, the approval flow and the isolation setup during the demo. Book a demo