Smart copy-paste: Jev knows what you copied and where it's going

Copy 94107 from a spreadsheet and paste it into a shipping form. It may be a ZIP code. Paste the same value into a CRM and it may be an account ID. A paste helper needs both sides of the action before it can offer a useful transformation.

Jev can classify the clipboard contents and the destination, then choose among a small set of formatting operations. Ordinary code performs the selected operation and preserves the original. The design below is a proposed workflow, with illustrative examples rather than a claim that a working clipboard product has been deployed.

Read the destination before changing the value

The destination often gives the strongest evidence. A field labeled “Postal code” tells you more than five digits alone. A Markdown editor, a spreadsheet range and a terminal prompt each accept different representations of the same copied text.

Collect the content type, the field label and the permitted operations. Where the application exposes an explicit schema, use it directly. A numeric field does not need a model to establish that it rejects a paragraph. A known JSON object can be parsed by a JSON parser.

Use Jev for the remaining ambiguity: whether a copied block is probably a contact card, whether a table belongs as rows or an attachment, or whether the destination appears to want plain text instead of rich formatting. TypeSafe's Choice primitive lets you define those options and inspect their probabilities.

Keep “paste unchanged” and “ask the user” available. Otherwise the model must choose a transformation even when every transformation would be wrong.

Make the transformation boring and exact

Separate the decision from the data operation. Jev can select plain_text, table_rows, markdown_link or unchanged. It should not invent the contents of the resulting table or reconstruct a phone number from memory.

A deterministic formatter can remove visual styles while retaining text, escape a label for Markdown, or parse tab-separated cells. Give every formatter a defined input, a defined output and a way to report that it could not safely transform the content.

Copied content	Destination	Proposed behavior
A URL and selected page title	Markdown note	Preview a link with the original URL
Three tab-separated columns	Spreadsheet range	Preview the cells before insertion
94107	Field with no useful label	Offer unchanged paste; do not infer an address
A formatted customer message	Plain-text support note	Remove presentation styling, retain exact words

Dates and numbers deserve particular care. Converting 03/04/2026 without a known locale can change the day. Removing the leading zero from an identifier can change the record. TypeSafe's documented limitations specifically recommend keeping numeric and date calculations in code. When the application lacks the information needed for an exact conversion, show the original value and ask.

A formatting result should carry the original clipboard revision, destination identifier and selected formatter. Recheck them immediately before insertion. If the user copies something else or changes fields while a decision is in flight, discard the stale result.

Keep the original one action away

Show a short preview of the proposed paste when the transformation changes structure. “Paste as three rows” is understandable. A silent rewrite is harder to inspect, especially if only one cell changed.

Let the user paste the original without visiting settings. Preserve enough local state for undo, including the content replaced in the destination. An undo action that restores the clipboard but leaves the wrong value in the form has not reversed the paste.

A terminal deserves an even narrower contract. The helper can show copied commands as text, with line breaks visible. Execution requires the user's separate action. Do not convert a classification such as “shell command” into permission to run it, and do not silently append a newline that submits it. Apply the same separation to form submission and sending a message.

This is the same division used in the support-agent budget workflow: a model may suggest an operation, while code and the user's granted permissions determine which operation is available.

Decide what can leave the device

Clipboard contents can include passwords, access tokens, private messages and customer records. Prefer local parsing and field metadata wherever those are enough. Exclude sensitive applications and known secret formats before any remote classification, and let users disable cloud assistance for a destination.

A secret detector will miss some secrets. Combine it with explicit user controls and a narrow set of participating contexts. Do not upload the clipboard merely because a paste event happened. If the helper cannot send the necessary context under the user's settings, ordinary paste should still work.

Jev consumes text, so an image or PDF needs a separate extraction step before semantic classification. That step has its own cost and data exposure. Offer it only when the user has asked for a transformation that needs it; an image pasted into an image field can stay an image.

Estimate a month from the pastes that reach cloud classification after local parsing and reusable cached decisions have handled the rest. Multiply that request count by the observed mean classification price, then add extraction and billable retries once each. Divide the resulting spend by all completed pastes to track cost per paste, including local ones. Keep that average separate from the price of a cloud request. A paste that works entirely on the device adds no model charge.

Test whether values survived

Build a small test set of paste pairs: leading-zero IDs, ambiguous dates, Unicode names, quoted CSV fields, multiline commands and tables containing formulas. Check the exact output bytes where exactness matters. Evaluate classification and formatting separately so a correct label cannot hide a broken parser.

Track how often users accept the preview, choose the original or undo immediately. Include cases where no transformation should happen. A helper that preserves 94107 when the destination is unclear has completed a useful decision without paying for a generative rewrite.

You can use Jev on Vaaya, along with 500+ model options and 1,500+ endpoints. One API key, one balance, a receipt for every call.

Questions

Does Jev rewrite clipboard data?

Jev selects among typed options or returns scores and probabilities. A separate formatter performs the transformation. Keep exact values, identifiers and arithmetic in ordinary code.

How should smart paste handle an ambiguous value?

Preserve the original and offer a choice. A value such as 94107 could be a ZIP code, an account identifier or a number; destination metadata can help, but the model's guess should not silently change it.

Can a paste helper execute a copied command?

Treat pasting and execution as separate actions. A classifier can choose a display or formatting mode, but it should not grant permission to run a terminal command or submit a form.