> Source: https://upvera.tech/research/an-ai-that-never-sees-the-number/
> Upvera Upi is the AI assistant members meet on Upvera; it runs Anthropic's Claude models and keeps platform numbers, publishing and UPV out of the model's reach.

UTR-26-08 · Research note · Rolling out

# An AI that never sees the number

Upvera Upi is the AI assistant members meet on Upvera; it runs Anthropic's Claude models and keeps platform numbers, publishing and UPV out of the model's reach.

## One door

Upi is the AI assistant a member meets on Upvera. A chat, help while writing a post, a draft built from recent commits: each is a surface of Upi, and each passes through the same server gateway. New assistant features arrive as surfaces or tools inside Upi, under the same checks and limits. Ranking is not part of Upi: Upvera Prism, which orders Discover and Home, is a separate system.

Upi runs on Anthropic's Claude models. A person picks a tier for each message, Fast, Balanced or Deep, from those their plan includes. Each step up uses a larger model with more room to think; Fast is the default. Upi never moves a conversation to a costlier tier on its own; at most, it suggests one.

Before the first turn, Upi asks for permission and says what is sent, who receives it (Anthropic) and what does not happen: Anthropic does not train models on this content, and passwords and payment details never reach Upi. Taking the permission back is as easy as giving it. Each turn carries the date and language as context, not a name or an email.

Upi's server has been live since October 10, 2026. It is rolling out to members gradually.

## Widgets are tools, numbers come from live data

A startup card, a comparison, a plan or a post draft in a Upi conversation is not text the model typed. It is a tool the model calls. The server checks the input against the tool's schema, checks every ID against the database (a startup must be visible to the person; a cited post must be live), and only then sends a block to the app. The block carries IDs and the model's words. The app draws the numbers from live data when it renders the card. A startup card shows today's Upvera Value because the app asks for it, not because the model wrote it.

Read tools apply the same rule from the other side. When Upi looks up a startup, it gets a description and summaries of recent posts, with no platform figures. Upi's tools never hand the model Upvera Value, rank, wallet balance or Prime counts, so it cannot round them, misremember them or make them up. The rule is enforced by what the tools hand over, not by a line in the prompt.

When a draft needs a figure, the model writes a sealed slot instead of a number. The tool fills it from the person's own record, for a startup they own, from a closed list of keys, and places the value only in the block that reaches the app. The value never travels back to the model. A sealed figure is fixed to the moment it was written, and a slot with no value stays visibly empty. Upvera Value and UPV are not on the list.

## A history that only grows, a cost held like money

A conversation is stored in the form the model provider expects, and it only grows. Messages are never edited, trimmed or removed from the middle. A conversation keeps the prompt and toolset version it started with; a change ships as a new version for new conversations.

Each turn carries an ID chosen by the app, so sending it twice brings back the saved result instead of a second answer. Content written by someone else, such as a post, reaches the model marked as untrusted data, never as instructions. The tools are narrow for the same reason: a model misled by what it reads still has no tool that publishes, messages or moves UPV.

Cost is handled like money. Before every model call, the server reserves the worst case for that call; afterwards it settles the real cost and releases the rest. The limits live in the database, not in code. If a reservation would cross a limit, the call does not happen. A call that breaks off midway counts at its full reservation, so nothing goes uncounted.

## What Upi will not do

| Action | Upi |
| --- | --- |
| Moves UPV | Never |
| Publishes a post | Always the person |
| Saves to memory | Always asks |
| Saves a private note to Hatch | Do, Ask or Never |

UPV is the unit members use to support startups; Upi never moves it. Upi never publishes, never follows anyone and never sends a message on its own. When an action has a side effect, Upi shows a confirmation card; the person taps, and the app calls the same path it would use without Upi.

One narrow class works differently: actions that are private and reversible. Its one member today: a note saved to Hatch, a founder's private notes. For this class, the person picks Do, Ask or Never; Ask is the default. On Do, the server writes the note, the database checks the rule again, and the conversation shows a receipt with Undo. Publishing, messages, memory and UPV can never enter this class.

## Upi suggests, a rule decides

The Seen seal marks a post whose stated change shows up in the product. Upi's job in it is small. It reads the founder's own post, and only the post, and suggests up to three checkable claims: a phrase on the site, a new page, a new style, a commit message or an App Store release note.

A deterministic matcher with no AI in it decides. It looks only at sources the founder has verified (a site, a repository, a store listing), in a fixed window around the post. Text on a page is not enough by itself; it has to arrive with a real release. When a claim meets a trace, the post gets the seal and a receipt naming the claim, the source and the moment. The receipt reads: Upi suggested the claim, a trace stamped it.

If Upi suggests nonsense, the result is no match. No seal is set, and no wrong seal is set. A claim that finds no trace is visible only to the founder. Upi never reads the pages being checked, so a page cannot send it instructions.

Behind this sits a wider rule: AI output never feeds the Upvera Value Engine. The engine's inputs must be reproducible, and a model's answer is not. Upi can show and suggest. It never decides a seal, a stage or a value.

## An agent may write; a person publishes

Upvera also runs a Model Context Protocol server, live since October 17, 2026. A person's own AI client, such as Claude, Cursor, VS Code or ChatGPT, can read that person's Upvera data and prepare posts. Here Upvera calls no model, so the number rule takes a different shape: the server gives numbers that are already public, such as a startup's Upvera Value with its latest receipt, and never personal figures such as a UPV balance. Sign-in uses OAuth 2.1 with PKCE; there is no API key.

The law is "An agent may write; a person publishes." In practice:

-   The assistant prepares a post. A card shows it exactly as it will appear, and the text is locked on Upvera's side.
-   The publish tool is offered only to the card, not to the model, and it needs a one-time ticket the model never receives.
-   Publishing is an optional permission the person ticks when connecting; a client cannot request it.
-   Pressing Publish sends a one-time ticket and a hash of the exact text, and Upvera runs the same checks the app runs. For 10 minutes, the card offers Undo.
-   A first post from an assistant, or a post with a link or an @mention, opens on Upvera's own compose screen to be confirmed there.
-   No tool moves UPV, reads direct messages or reads anyone else's private data.

Publishing from the card is rolling out. Every connection is listed in the person's settings and can be revoked at once.

## References

-   Upvera MCP server: [https://upvera.dev/mcp/](https://upvera.dev/mcp/)
-   MCP tools reference: [https://upvera.dev/mcp/tools/](https://upvera.dev/mcp/tools/)
-   Security at Upvera for developers: [https://upvera.dev/security/](https://upvera.dev/security/)
-   Model Context Protocol: [https://modelcontextprotocol.io](https://modelcontextprotocol.io)
-   MCP Apps extension: [https://github.com/modelcontextprotocol/ext-apps](https://github.com/modelcontextprotocol/ext-apps)
-   Upvera: [https://upvera.co](https://upvera.co)

[Next · UTR-26-09**Upvera Prism: learning only from people**](/research/prism-learning-only-from-people/)