Aura Wire Chat: live cost, fewer tokens, secret guard, local routing

See tokens and estimated cost on every turn in VS Code.
Route cheaper, compress repeats, chat with Copilot, Cursor, Claude, OpenAI, Gemini, or Ollama.

Product page
The optional Windows desktop app adds the full command center and the wire CLI.

Screens on this page use sample data. Token and cost figures are illustrative.

Install

1. Install the extension and open a chat from the Aura Wire view. Each chat is an editor tab you can split, drag, or move to another window.
2. The first chat downloads and starts Aura Wire Host in the editor. No certificate, proxy setting, or admin prompt.
3. Pick a destination and a model: Cursor, Copilot, Claude, OpenAI, Anthropic, Gemini, Ollama, or any other configured provider. Starred models stay at the top.
4. Add keys with Aura Wire: Provider keys, from the Command Palette or Provider keys at the bottom of the model picker. Host stores them on this machine. If a provider is missing, the picker opens that same screen.

Host starts with a 14-day full-pipeline trial (no card). When it ends, the whole product turns off until you buy the one-time Aura Wire Chat license; the product page has the details. Works in VS Code 1.85 and newer, and in Cursor and Windsurf.

How this chat is different

Other chats hide the request. This one shows it. Under every reply: which model answered, tokens in and out, how many Compress removed, and the estimated cost. The composer shows a running total and the context budget. Long chats keep the full transcript; only the outgoing request is trimmed, oldest tool results first.

The pipeline between the composer and the provider is yours to set. The wrench on the composer turns Guard, Router, Compress, and Advise on for this chat, and opens the routing profile that chat is using.

Guard strips secrets before the request leaves

An API key went into the composer. Guard redacted it before the request left the machine, and the footer says so.

Router sends the turn to a cheaper model

One rule: OpenAI turns scored low complexity go to gpt-4o-mini; everything else falls through unchanged. Rules run top to bottom, first match wins.

Rules match on provider, model, token range, attachments, tools, project, or time of day, and can change the model, route to another API, or delegate to a local agent. When a rule fires, the reply footer names the model that answered instead of the one you asked for, with the tokens saved. Open the designer from the wrench on the composer, or run Test against recent traffic to see which rules would have matched.

Compress trims the request before it leaves

Tokens saved per provider, in tokens and in dollars, with the rate chart for the last hour.

Compress drops repeated context and stale tool results from the outgoing request; the transcript in the tab is never touched. The footer under each reply shows how many tokens it removed, and this board adds them up by provider.

Cost and traffic, in the editor

Where this billing cycle is heading, split by project and by provider.

Aura Wire boards open as editor tabs: cost projections, live traffic, Router and Compress stats, and Guard activity. Observe is on from the start; Guard, Router, and Compress stay off until you turn them on from Pipeline overview or the wrench on a chat.

It asks before it writes or runs

The model asked to run node --test. The extension stopped and asked first.

Writes and shell commands wait for run once, skip, or Always. Reads run on their own. Each turn lists the files it touched, with Keep / Undo per file; Undo shows what it will restore. Edits go through the editor, so Ctrl+Z works on them too.

Cursor, Copilot, and Claude run their own agent inside Host, so those cards do not apply. Their tool calls still show in the transcript as read-only cards (needs a Host that forwards nested_tool for that destination).

What this is, and what it is not

Frequently asked questions

Does this replace Copilot Chat or Cursor Agent? No. It is a second chat that can route to them, and to other providers, through Host. Many people keep both open. Aura Wire is the one that shows where the request went and what it cost.

Can I use a local model? Yes. Ollama is a destination like any other, so a chat can move from a hosted model to a local one mid-thread. Host runs on loopback either way.

Where do prompts and API keys go? Host listens on loopback. You enter keys on Provider keys, or set the vendors' usual environment variables; the extension never stores them. Prompts and responses stay in memory and are not saved. Host keeps sanitized metadata: provider, model, token counts, latency, and cost. Only requests sent through this extension, the wire CLI, or an OpenAI-compatible client are observed. Capture is not enabled, so there is no driver, certificate, or administrator step.