Velden is an AI agent built by Tendery UG (haftungsbeschränkt). It works for the organization that installed it and writes its Slack messages, pull requests and commits with AI models. This page says how Velden marks what it writes, so that people and software can tell it was generated by AI (Art. 50 of the EU AI Act), and where a mark is missing or gets lost.
The visible label
Velden's first message in each Slack thread ends with a line that says it is an AI agent, in the language Velden writes in there. In English:
Velden is an AI agent working for this workspace. It can be wrong, so check what matters before you rely on it.
In German:
Velden ist ein KI-Agent, der für diesen Workspace arbeitet. Er kann sich irren, prüfen Sie also Wichtiges, bevor Sie sich darauf verlassen.
The thread's language is the one Velden replied in during its conversation turns there. A thread where that is German gets the German line, and every other thread gets the English one, including a thread where Velden has not replied in a conversation turn yet. The first line a thread shows sets its language, and every later line there uses it.
The line also shows:
- on every message Velden starts itself, in a channel or as a direct message (the welcome message to the person who installed Velden is in English),
- on the next message in a thread after someone Velden has not heard there before writes in it,
- in a thread that has not shown it for 30 days,
- on the messages where Velden reports or asks someone to approve: the message at the end of a run, the message asking an owner for a decision, and the confirmations that ask someone to approve a merge or a setting.
Velden's other messages in a thread carry the machine-readable mark below and no visible line.
Every pull request Velden opens says at the top of its description, in both languages, that it comes from Velden, an AI agent.
Machine-readable marks
Slack messages. Each message Velden posts or edits carries Slack message metadata with the event type velden_ai_generated and this payload:
{
"schema": 1,
"ai_generated": true,
"system": "Velden",
"provider": "Tendery UG (haftungsbeschränkt)",
"info": "https://velden.ai/en/legal/ai-transparency"
}
The payload is the same for every message and holds no task, person or text. Any member or app of the workspace can read it: conversations.history and conversations.replies return it with include_all_metadata=true, and an app can subscribe to the message_metadata_posted event for this event type. The message is also posted by Velden's Slack app, which Slack shows as the sender.
Pull requests. The first line of each description is an HTML comment that GitHub does not show and its API returns in the body field. It holds velden:ai-generated schema=1 info=https://velden.ai/en/legal/ai-transparency. The author of the pull request is Velden's GitHub App, which GitHub reports as a bot.
Commits. Each commit with a change Velden wrote, and the merge commits Velden makes, end with these trailers:
Generated-by: Velden (Tendery UG)
Velden-Run: <the id of the run>
git log --format='%(trailers:key=Generated-by)' lists them. Merge commits and reverts belong to no run and have only the first trailer. GitHub signs the commits Velden's App makes through the API, so they show as verified.
The schema number changes when a mark's fields change, and the event type and marker name stay the same.
Where a mark is missing or gets lost
- Velden's later messages in a thread carry the metadata and no visible label, so someone who sees only a later message (in a notification or a search result) sees no label.
- Messages only one person sees (ephemeral messages) carry no metadata, because Slack keeps none on them. They carry the label in a thread that has not shown it yet, and always on the confirmations above. A reply Slack delivers through a command's response URL carries no metadata either. A new one carries the label in English. When someone presses the button on a confirmation, the message that replaces it carries the label in the thread's language, and any other replacement keeps the label of the message it replaces.
- The short replies Slack shows right after a slash command (a usage hint, or "Got it.") carry neither the label nor the metadata.
- A card replaced after a button press may lose its metadata. We have not verified yet whether Slack keeps it.
- A file Velden uploads, such as a walkthrough video, has no metadata, because Slack's file upload takes none. Its comment carries the label when the thread has not shown it yet.
- Slack notifications show a message's text without the label.
- Text copied out of Slack or GitHub carries neither the label nor the metadata.
- A person who edits a pull request description can remove its marker. Velden's own later edits keep it.
- A squash merge made by a person may drop the trailers. The pull request keeps its marker and its author.
- Watermarks are unreliable or absent on short text (under about 200 tokens), which includes most Slack replies.
- Velden marks no file it changes. The European Commission's guidelines on Art. 50 treat source code, configuration and code comments as outside the marking duty, and a marker inside a file would change the code people reviewed.
Watermarks from the model providers
Velden adds no watermark of its own to its text. What the model providers say, as checked on 28 October 2026:
- Anthropic states that text from current Claude models carries a watermark, including on Google Cloud, and lists Claude Sonnet 5.5, Opus 5.5 and Haiku 5.5, which Velden uses (announcement, help article). Anthropic has not stated whether this covers Vertex AI's EU multi-region, which Velden uses. Its detector is a preview that Anthropic gives access to on request.
- Google has published no statement that text from Gemini models on Vertex AI carries a watermark. Most of Velden's Slack messages are written by Gemini, so for them the label and the Slack metadata above are the marks we can show. Google's documentation of the voice model that narrates walkthrough videos does not mention a watermark either. The video's comment says an AI voice narrates it.
- Review findings in a pull request can come from an OpenAI model through OpenRouter. OpenAI's EU text watermark is opt-in on OpenAI's own API (TechCrunch, 5 October 2026), so we know of none for that route.
We update this page when a provider publishes more.
Questions
Questions about these marks or how to detect them go to privacy@tendery.ai.