Coding Agents

One npm install — compress LLM context for every major coding agent. Auto-detects Cursor, Windsurf, Continue, Cline, Claude Code, Codex, and Aider. Track savings in your dashboard.

How it works

SuperCompress runs a local HTTP server that sits between your coding agent and its LLM provider. Every prompt your coding agent sends is intercepted, compressed by the SuperCompress engine, and then forwarded to the LLM. The response streams back to your agent untouched.

# Without proxy
Coding agent ──→ long prompt ──→ OpenAI / Anthropic
                                    $$$ per token

# With SuperCompress
Coding agent ──→ long prompt ──→ [ SuperCompress ] ──→ compressed ──→ OpenAI / Anthropic
                                    cuts ~65%                              $$

The server runs on localhost:8080 by default. It uses the SuperCompress compiler mode to maximize token savings while preserving answer-critical evidence. Every compression is tracked and reported to your dashboard.

Installation

Install SuperCompress globally via npm:

npm install -g supercompress

Verify the install:

supercompress --version
# → 1.0.0

Setup (auto-detect agents)

Run the setup command to auto-detect your installed coding agents:

supercompress setup

Setup scans your system for:

• Cursor — detects config in ~/.cursor/
• Windsurf — detects config in ~/.codeium/
• Continue — detects config in ~/.continue/
• Cline — detects config in VS Code settings
• Claude Code — detects claude CLI in PATH
• Codex — detects codex CLI in PATH
• Aider — detects aider in PATH and .aider.conf.yml

For each detected agent, supercompress asks if you want to wire it up. Say yes, and it writes the config to point to localhost:8080.

$ supercompress setup
🔍 Scanning for installed coding agents...
✓ Cursor found (~/.cursor/config.json)
✓ Continue found (~/.continue/config.json)
✓ Claude Code found (/usr/local/bin/claude)
? Wire up Cursor to use SuperCompress? [Y/n] Y
✓ Cursor configured → http://localhost:8080/v1/chat/completions
? Wire up Continue to use SuperCompress? [Y/n] Y
✓ Continue configured
? Wire up Claude Code to use SuperCompress? [Y/n] Y
✓ Claude Code configured

Setup complete! Run supercompress start to start.

Start the server

supercompress start

This starts the server on localhost:8080. Keep it running in the background while you code. The server:

• Listens for requests from your coding agents
• Compresses each prompt using the SuperCompress compiler
• Forwards the compressed prompt to OpenAI or Anthropic
• Streams responses back to your agent
• Logs compression stats to the console
• Sends usage data to your dashboard every 60 seconds

$ supercompress start
┌──────────────────────────────────────────────┐
│  SuperCompress running on              │
│  http://localhost:8080                        │
└──────────────────────────────────────────────┘

✓ Detected agent: Cursor
✓ Detected agent: Continue
✓ Detected agent: Claude Code

[15:42:01] COMPRESS | cursor | 2,341 → 853 tok | -63.6% | risk:low
[15:42:08] COMPRESS | continue | 4,120 → 1,482 tok | -64.0% | risk:low
[15:42:15] COMPRESS | claude-code | 1,880 → 674 tok | -64.1% | risk:low

Supported agents

Agent	Auto-detect	Detection method
Cursor	✅	Reads ~/.cursor/config.json
Windsurf	✅	Reads ~/.codeium/ config
Continue	✅	Reads ~/.continue/config.json
Cline	✅	Scans VS Code settings
Claude Code	✅	Checks claude in PATH
Codex	✅	Checks codex in PATH
Aider	✅	Checks aider in PATH + config file

Dashboard tracking

Once the server is running and your agents start making requests, visit your dashboard to see:

• Coding Agents tab — per-agent stats with requests, tokens in/out/saved, savings %
• Summary cards — total agents connected, total requests, total tokens saved
• Per-agent table — last used timestamp, agent icon, and individual performance
• CTA banner — if no agents are configured, shows install instructions

The server sends usage data to your account every 60 seconds. No prompt content is sent — only token counts, agent name, and timestamps.

Streaming support

The server supports streaming responses from both OpenAI and Anthropic. When a coding agent sends a streaming request, the server:

1. Compresses the prompt (non-streaming)
2. Forwards the compressed prompt with stream: true
3. Pipes each SSE chunk back to the agent in real time

Streaming works transparently — your agent doesn't know the server is there. No extra latency beyond the initial compression (~60ms).

Manual configuration

You can also configure agents manually by pointing them to the server endpoint:

Provider	Endpoint
OpenAI-compatible	http://localhost:8080/v1/chat/completions
Anthropic-compatible	http://localhost:8080/v1/messages

Set the API key to your actual OpenAI or Anthropic key in the agent's config. The server reads OPENAI_API_KEY and ANTHROPIC_API_KEY from the environment and passes them through.

Uninstall

To remove SuperCompress and restore your agents to their original configuration:

supercompress setup --reset
npm uninstall -g supercompress

The --reset flag restores each agent's config to its original state (using the backup created during setup).