GlyphCompress — Cut Your LLM Token Bill
Your IDE re-sends the same open files to the model on every turn, and you pay for them every time.
GlyphCompress cuts what a coding session is actually billed — measured against real provider tokenizers (OpenAI's cl100k_base/o200k_base), never a character count. It works on four fronts, and the two largest compress nothing at all: it stops re-transmitting files the model already has (−78.5% billed over ten turns), places provider cache breakpoints where they genuinely cover the request, compresses the content that does need sending (26% aggregate), and condenses old turns once they stop earning their keep.
We measured the glyph encoding this project is named after, and it lost. Against real tokenizers it costs tokens on real files instead of saving them, so it is off by default — held there by a rule that refuses any transformation that would send more tokens than it received.
Don't take those numbers on faith. Get your own, from your own code:
npx glyph-compress measure src/your-biggest-file.ts --turns 10
What this extension does
- Compress Selection: compress the currently selected text/code and copy the result to your clipboard.
- Ask LLM (Auto-Compress): automatically compress context before it reaches VS Code's built-in Language Model API (Copilot Chat and compatible extensions), with zero manual steps.
- Zero-Command Transparent Proxy: a local proxy that sits between your IDE (Cursor, Continue, Cline, and other OpenAI/Anthropic/Gemini-compatible tools) and the real provider API, compressing every request automatically. Point your IDE's API base URL at
http://localhost:8080, then see the two setup notes below.
- Build Project Codebook: scan your workspace to rank files by relevance to a task and build a persistent index used by the Context Router.
- Compress Entire Workspace: batch-compress every open/relevant file at once.
- Show Compression Stats / status bar: live token-savings numbers for your current session.
Using the proxy: two things that are easy to miss
Start it first. Nothing is compressed until the proxy is listening, and clients do not say so helpfully — Continue reports a bare Connection error. Run GlyphCompress: Start Zero-Command Proxy from the Command Palette, then check:
curl -s -o /dev/null -w "%{http_code}\n" http://localhost:8080/stats
200 means it is up; 000 means nothing is listening. Starting it from the palette (rather than a terminal) is also what makes Show Compression Stats move — a terminal-started proxy keeps a separate counter and reports to http://localhost:8080/dashboard instead.
One proxy forwards to one upstream. glyphCompress.targetApiUrl picks it. If your client lists both an OpenAI model and a Gemini model against localhost:8080, only the family matching the current target will work, and the other fails with an upstream error that looks like a credential problem. Your client's own provider field should stay openai regardless — it describes the format spoken to the proxy, not the company at the far end.
Providers and trust policies
Works with OpenAI, Anthropic Claude, Gemini, and Antigravity — each has its own compression profile, calibrated separately since tokenizers behave differently across providers.
Trust policies control what compression is allowed to do to your data:
| Policy |
Reversible |
What it allows |
lossless |
Yes |
No lossy transforms at all |
reversible |
Yes |
Prompt patterns, tech names, file paths, diagnostics, dynamic dictionary |
privacy |
Yes |
Everything in reversible, plus redaction of API keys, tokens, emails, and IPs before compression |
lossy |
No |
Everything, including code minification, structural summaries, and redundancy stripping |
When a level/policy combination allows something irreversible (like ultra-level code summarization), the extension surfaces a plain-language trust warning in its output channel — derived directly from what that policy actually permits, not a guess about how the model will behave.
Settings
| Setting |
Default |
Purpose |
glyphCompress.enabled |
true |
Enable/disable compression |
glyphCompress.provider |
auto |
auto, raw, openai, anthropic, antigravity, gemini, local |
glyphCompress.compressionLevel |
standard |
light, standard, aggressive, ultra, auto (auto picks per request) |
glyphCompress.trustPolicy |
auto |
lossless, reversible, privacy, lossy |
glyphCompress.targetApiUrl |
https://api.openai.com |
Upstream API the proxy forwards to |
glyphCompress.showStatusBar |
true |
Show live compression stats in the status bar |
glyphCompress.experimentalDecay |
false |
Progressively compact older chat history for long-running conversations |
glyphCompress.holographicFolding |
false |
Group overlapping related files and shared imports into layered blocks |
glyphCompress.intentDiffs |
false |
Condense verbose diffs into short, declarative change-action lines |
Also available as
- CLI:
npx glyph-compress <file> — same engine, usable outside VS Code, plus route/team-codebook/inspect/doctor commands.
- MCP server:
npx glyph-compress-mcp — use GlyphCompress from Claude Code, Claude Desktop, or any other MCP-compatible client.
- Library:
npm install glyph-compress for direct GlyphCompressor/wrapOpenAI/wrapAnthropic usage in your own code.
Links
Dual-licensed: AGPL-3.0 for open-source use, commercial license available for proprietary use.
