Prompt Proxy for VS Code

A local prompt optimizer, semantic cache, and conversational AI agent — all running inside VS Code without leaving your editor.

🔒 Fully local — zero telemetry. All prompt processing, token counting, semantic caching, and secret scanning run entirely on your machine. No prompts, code, file contents, or metrics are ever sent to any external server. The only network request is the final Copilot call you explicitly trigger.

What it does

Prompt Proxy sits between you and Copilot. Before your prompt reaches the model it:

1. Compresses filler words, softeners, and redundant phrasing
2. Checks the local semantic cache — reuses a prior result if a similar prompt was already answered
3. Packs workspace context — active file, open editors, diagnostics, recent saves
4. Estimates token cost before the request is sent
5. Calls Copilot and streams the answer directly in the sidebar (Agent mode) or hands you the optimized prompt to review first (Optimize mode)
6. Remembers the conversation across turns per workspace so follow-up pronouns ("fix it", "add tests for that") resolve correctly

Modes

Select the mode from the Mode dropdown in the sidebar or via the status bar item ($(robot) Proxy [Agent]).

Switch mode any time by:

• Clicking the status bar label ($(robot) Proxy [Agent]) → QuickPick
• Changing the Mode dropdown in the sidebar
• Typing @promptproxy /mode agent (or optimize / direct) in Chat

Sidebar panel — Prompt Proxy Control

Open it from the Chat sidebar or press the status bar item.

┌──────────────────────────────────┐
│  Prompt Proxy       Local cache  │
│  Mode  [ Agent ▾ ]               │
│ ┌──────────────────────────────┐ │
│ │ Type your prompt here…      ↑│ │  ← send button (icon, like Copilot)
│ └──────────────────────────────┘ │
│  ⚠ Alerts (secrets, errors)      │
│  [ Open @promptproxy ]           │
│  [ Use optimized ] [Copy] [Docs] │
│                                  │
│  ┌── Analysis table ──────────┐  │
│  │ Cost │ $x  │ Saved │ n tok │  │
│  │ Tok  │ n→m │ Est.  │ n out │  │
│  │ Cache│ semantic 82%        │  │
│  │ Price│ in $x + out $y      │  │
│  └─────────────────────────────┘ │
│  Refinements │ Optimized prompt  │
│  • use nouns │ [compressed text] │
│                                  │
│  ┌── Copilot response ────────┐  │
│  │ streaming…                  │  │
│  └─────────────────────────────┘ │
└──────────────────────────────────┘

Send button

The circular ↑ button inside the textarea behaves like the Copilot send button — hover shows the current mode action ("Run Agent — optimize + call Copilot").

Chat participant — @promptproxy

Type in the VS Code Chat panel:

@promptproxy refactor the auth middleware to use async/await

Every message is automatically optimized. Copilot's answer streams back in chat.

Commands

Conversation memory

Up to 12 turns per workspace are remembered. Back-references resolve automatically:

Turn 1: "refactor the auth middleware to async/await"
Turn 2: "now add unit tests for it"
         ↑ proxy injects "[Continuing from: 'refactor…']" before sending

Analysis table (Optimize mode)

After analyzing a prompt, the result card shows a compact table:

Below the table, Refinements and the Optimized prompt sit side by side.

Secret detection

If your prompt contains what looks like an API key, token, or private key header, a ⚠️ alert appears in the panel before anything is sent. Patterns detected:

• OpenAI keys (sk-…)
• Anthropic keys (sk-ant-…)
• GitHub tokens (ghp_…, ghs_…)
• AWS access keys (AKIA…)
• PEM private key headers
• Generic password=, token=, api_key= assignments

Semantic cache

The local SQLite cache stores every prompt you send and builds embeddings for semantic similarity. On future prompts it checks for:

• Exact match — returns the cached optimized version instantly
• Semantic match (≥ 68% cosine similarity) — returns and boosts confidence score
• Miss — optimizes fresh, writes to cache

The cache is seeded on activation from your git log, Copilot chat history, README, package.json, and AI instruction files (.github/copilot-instructions.md, AGENTS.md, etc.).

Manage via command palette:

• Prompt Proxy: Show Cache Statistics
• Prompt Proxy: Clear Semantic Cache

Privacy & Security

To verify: open the extension source (out/extension.js) or the engine source (engine/dist/). Search for http, fetch, axios, request — you will find zero outbound calls outside of the Copilot API.

Command palette

Configuration