foveance-proxy (npm launcher)

Cut LLM input tokens up to 82% — losslessly. Not one line of your app changed.

Your AI agent re-sends its whole history on every turn — the same listings, tool outputs and stack traces, over and over. Foveance is a real lossless codec for that context: like gzip finds repeated bytes, it replaces text that already appeared with a compact back-reference, removing tokens without removing information. Nothing is dropped, every fact survives.

This is a tiny Node launcher for the Foveance proxy, so users of Node-based AI tools (Claude Code, OpenAI Codex, opencode, Crush, Continue, ...) can start it with one command, without a manual Python step. The proxy is OpenAI- and Anthropic-compatible, streams, and forwards your API key untouched — you only point a client's base URL at it.

Prerequisites

• Node.js >= 16
• Python >= 3.10 with Foveance installed once:

pip install foveance

Use

# OpenAI upstream
npx foveance-proxy --upstream https://api.openai.com/v1

# Anthropic upstream (Claude Code, Anthropic SDK)
npx foveance-proxy --upstream https://api.anthropic.com/v1

# Local Ollama / vLLM / TGI / LM Studio
npx foveance-proxy --upstream http://localhost:11434/v1

All flags are forwarded to foveance proxy (--host, --port, --budget, --drift, --policy, --upstream, --cache-aware, --price-per-mtok), and the FOVEANCE_UPSTREAM / FOVEANCE_BUDGET / FOVEANCE_DRIFT / FOVEANCE_POLICY environment variables are honoured too. A live tokens-saved dashboard serves at http://localhost:8799/ while the proxy runs.

Even simpler: the Python CLI has foveance wrap claude (or foveance wrap -- <any command>), which starts the proxy, routes the tool through it, and prints a tokens-saved summary on exit.

Then point your tool at it (one variable):

# OpenAI-style tools, Codex, Ollama-backed clients
export OPENAI_BASE_URL=http://localhost:8799/v1

# Anthropic-style tools, Claude Code
export ANTHROPIC_BASE_URL=http://localhost:8799

See the main README and docs/usage.md for per-tool recipes. This launcher is a convenience wrapper; the proxy itself lives in the Python package.

Apache-2.0.