Web Clipper: AI Interpreter

The interpreter lets a template ask a language model for something the page does not state outright.

Prompt variables

Write a natural-language prompt directly in a template:

{{"a summary of the page"}}
{{"the three main arguments, as a bulleted list"}}
{{"is this technical or introductory? one word"}}

The model reads the page and fills in the answer as the clip is created.

Bring your own provider and key

The interpreter uses your model and your API key. Supported providers:

• Anthropic Claude
• OpenAI
• Google Gemini
• DeepSeek
• Perplexity
• OpenRouter
• Azure OpenAI
• Hugging Face
• Meta
• Ollama — locally hosted models

You can also add custom providers and models.

The Ollama option is worth noting: with a local model, page content never leaves your machine at all.

Privacy

Interpreter requests go directly from your browser to the provider you chose. MDflow does not gather or store request data.

MDflow is not a middleman here. It does not see the page content you send, the prompts, or the responses. Your API key is yours, your bill is yours, and your data goes only where you pointed it.

The corollary: the page content is sent to whichever provider you configured, under that provider's terms. If you clip something confidential, that matters — use a local Ollama model, or turn the interpreter off for that clip.

Settings

• Run automatically when a template uses prompt variables, or trigger it manually.
• Override the interpreter context per template — give the model different instructions for a research paper than for a recipe.
• A live token usage counter shows what you are spending as you interpret.

When to use it — and when not to

Good uses, where the answer genuinely is not in the page:

• A summary in your own preferred length.
• A categorization — "which of my folders does this belong in?"
• Extracting an argument's structure.
• Translating.
• Reformatting messy content into a consistent shape.

Poor uses, where you are paying a model to do what a variable already does perfectly:

• The title, author, or published date — use preset or meta variables.
• Structured data the page already declares — use schema.org variables.
• Anything a CSS selector reliably gets.

See 06 Variables and Filters. The rule of thumb: if the page states it, extract it. If the page implies it, interpret it. A model asked for the title will usually be right, occasionally be creative, and always cost more than reading the <title> tag.

The payoff

The combination that makes this worthwhile: a template triggered on a site you read often, that clips the article, extracts its declared metadata into properties, asks a model for a three-line summary and a suggested category, and files it — all from one keyboard shortcut.

The result lands in your MDflow workspace, where your other AI agents can read it through the API or MCP server. The clipper is where content enters that loop.