Build models in Claude or ChatGPT

Describe a calculation, quote, budget or rule set in a chat. The assistant writes a Valem model and the chat shows it as a live card: change an input and every value recomputes, with the numbers coming from Valem’s engine rather than from the assistant.

Add the connector
Ask for a model
Published calculators
Share or embed it
/mcp/app or /mcp?
Limits on the public sandbox

Add the connector

The sandbox hosts the chat card server at https://valem.run/mcp/app. There is nothing to install and no account to create.

Claude (web or desktop app): Settings → Connectors → Add custom connector, with the URL https://valem.run/mcp/app.
ChatGPT: turn on developer mode (Settings → Apps & Connectors → Advanced settings), then create a connector with the same URL.
Any other MCP client connects to the same URL over Streamable HTTP. Clients that support MCP Apps show the card; the rest get the computed values and a link as text.

Ask for a model

Ask in plain words, for example “Make me a quote calculator for a cleaning business: price per room, a weekend surcharge and a minimum charge.” The assistant:

reads valem_spec_guide once, to learn the spec format;
writes a ModelSpec with a view and a few embedded tests, and calls render_model;
fixes the spec and tries again if Valem reports errors or a failing test. A spec whose own tests fail is never shown.

The card appears in the conversation. Edit any input and it recomputes in place; a value that breaks a hard rule is put back, with the reason shown. The assistant is told the new values after each edit, so you can keep asking questions about them.

Why is this number? Hover a computed value to see its formula, the values it used and the inputs it comes from. Ask the assistant the same question and it calls explain_value, so its answer quotes Valem’s formula and values instead of a guess.
Several views. A model with more than one view gets tabs, and where the chat app allows it a Full screen button for bigger models.

Published calculators

For net salary, car tax, self-employed income and similar questions, the assistant can show one of the published calculators instead of writing a model: list_calculators finds it and show_calculator puts it in the chat as the same live card, with its tax year and cited sources. Ask, for example, “Show me Germany’s net salary for 60,000 a year.”

Each card comes with two links:

a share link (https://valem.run/#share=1.…) that opens the model in the sandbox, with the values you entered;
an embed URL (https://valem.run/embed.html#e=1.…) to put in an <iframe> on any web page.

The card’s Open in browser button opens the share link with the values on screen right now. The whole model rides in the link, so nothing has to be saved anywhere for it to work. A very large model can be too long for a link; then there is none.

`/mcp/app` or `/mcp`?

The sandbox runs two MCP servers for two different jobs.

	`/mcp/app` (this page)	`/mcp` (Connect your agent)
Where the model lives	In the chat, as a card	In a session on valem.run, shown in a browser tab
Setup	Add the URL	Add the URL, then pair the agent with a browser tab
State	None kept: each edit is recomputed from the spec the card holds	Kept in the paired session; you and the agent edit the same model
Tools	`render_model`, `explain_value`, `show_calculator`, `list_calculators`, `valem_spec_guide`	The full model tool set: create, mutate, evolve, explain, …
Best for	Chat users who want a working calculator in the conversation	Developers iterating on a model with an agent

Both are free to try on the public sandbox and need no Java or download.

Limits on the public sandbox

The hosted server allows 60 requests a minute per IP, and 20 new or changed models an hour. A model the server hasn’t seen for an hour is forgotten; the card sends its spec again on the next edit, so you won’t notice. Treat anything you put in a public sandbox as public.

Full tool detail, including the card’s result shape: MCP tools reference.