supercompress-proxy

SuperCompress

SuperCompress v2 — cut ~64% of LLM input tokens for coding agents, without losing the answer.

~400M Neural Keep on the SuperCompress API scores bulky context (files, logs, tool dumps, pastes) against the current question and drops the rest. Your ask stays intact. This package wires Cursor / Claude Code / Codex / 40+ harnesses into that API via MCP + hooks.

Website · Benchmarks · Playground · Docs

Install

npm install -g supercompress-proxy

Requires Node.js 18+.

Quick start (recommended)

One command links your account and auto-adds MCP + hooks across 40+ harnesses (25+ get a native auto MCP plugin — Cursor, Claude Code, Codex, Goose, Zed, OpenCode, fx, Hermes, OpenClaw, Grok Build, Grok Bot (hosted MCP URL), Gemini, Windsurf, Continue, and more):

supercompress setup
supercompress doctor

Then restart your agent so integrations reload. That’s it.

Re-detect later (new agent installed, etc.):

supercompress plugin
supercompress agents
supercompress doctor

Any other agent (custom harness, closed-source, DIY):

supercompress agents connect

That prints a stdio MCP snippet and writes an Agent Plugins 1.0 pack you can drop into any compatible client.

Benchmarks

Same keep-budget (35% of tokens kept). Who still has the answer?

Method	Answer-critical kept
FIFO / truncation	24.8%
Summarization	60.5%
H2O	97.9%
SuperCompress	100%

Metric	Result
Oracle recall (fixed budget)	100%
Mean token cut (real suite)	~67%
Important lines kept (compiler)	100%

Full methodology and charts: supercompress.dev/benchmarks

How it works

Your agent ──→ SuperCompress (hooks / MCP) ──→ smaller context ──→ model
                      ↑
                 query stays whole; only context is compressed

• You ask a question (never rewritten).
• Large context is scored against that question.
• Evidence-critical lines stay in original wording; filler drops.
• You pay for fewer input tokens.

Commands

Command	What it does
supercompress / tui	Interactive paper-branded UI (default in a TTY; Bun)
supercompress setup	Recommended — link account, detect agents, install MCP + hooks
supercompress plugin	Refresh agent integrations anytime
supercompress doctor	Per-harness health matrix (account / MCP / hooks)
supercompress agents	40+ catalog — auto-plugin vs recipe/connect
supercompress start / stop / status	Optional local proxy (setup --proxy)
supercompress usage	Plan, quota, savings (--json ok)
supercompress limits	Provider rate-limit runway — 5h / weekly windows (--json for agents)
supercompress uninstall	Remove configs under ~/.supercompress

Optional localhost API proxy (base-URL rewrite) if you explicitly need it:

supercompress setup --proxy
supercompress start

Then point OpenAI/Anthropic-compatible clients at http://localhost:8080/v1.

MCP

setup / plugin registers the MCP server on every detected host. You can also run it directly:

supercompress-mcp

Manual registration:

{
  "mcpServers": {
    "supercompress": {
      "command": "supercompress-mcp"
    }
  }
}

Tool	Purpose
compress_context	Compress bulky context for a query
connect_account	Link this install to your dashboard
usage_summary	Savings for the connected account
runway_check	Provider rate-limit runway (local, no account needed)

Runway (rate limits)

supercompress limits --json / MCP runway_check tell the agent how much provider quota is left and what to do with it. Built from limitcheck, extended to every harness SuperCompress plugs into. Runway runs locally, inside this package; nothing goes through the hosted API, and the hosted MCP has no runway tool. setup / plugin install it automatically: they add the runway rule to every agent's instructions and skill, register runway_check on the local MCP, copy a standalone runtime to ~/.supercompress/runway/, and connect Claude Code's statusline. uninstall reverses all of it.

Source	Harnesses
~/.codex/sessions rate-limit events (zero config)	Codex
Statusline tee, wired by setup (or supercompress limits connect claude-code). Your own statusline keeps working unchanged	Claude Code
Upstream rate-limit headers seen by the local proxy (Anthropic, OpenAI, xAI, OpenRouter, Codex)	Any harness routed through localhost:8080 (Aider, Continue, Cline, OpenCode, Goose, …)
supercompress limits ingest --provider <id> (stdin JSON) / SUPERCOMPRESS_LIMITS_SNAPSHOT / LIMITCHECK_SNAPSHOT	Anything else (Cursor, Windsurf, Grok Build, Gemini CLI, custom hooks)

Actions: go (with scope), land (keep working, checkpoint before each step), pace (balanced policy only), wait (actually exhausted — waitMinutes, alternatives), unknown (no source; never invented).

The default max policy uses a 5h window all the way to the provider's cutoff instead of stopping at a fixed 10% floor: short windows reset on their own, so unused quota is wasted. Weekly/monthly windows are judged by pace (remaining % vs time left). --policy conservative keeps limitcheck's 25% / 10% thresholds. In a seeded simulation (node test/runway-sim.js) max uses 100% of a binding 5h window vs 91.5%, for +7.8% completed steps, and checkpoints before 59 of 60 cutoffs. SUPERCOMPRESS_RUNWAY=0 stops setup from touching the statusline and disables proxy header capture. Only rate-limit numbers are stored (~/.supercompress/limits/), never prompts or responses.

Account & pricing

Launch promo: 5M tokens/month free, then $0.10 / 1M from the dashboard.

Privacy

Hooks / MCP run on your machine. Provider API keys stay with your agent. Context text is sent to the SuperCompress API so the hosted compiler can compress it.

More

• Coding agents: https://docs.supercompress.dev/coding-agents
• HTTP / Python API: https://docs.supercompress.dev/quickstart
• Source: https://github.com/Supercompress/Supercompress

License

MIT — see LICENSE.