tare

# tare

> Lossless-by-default, cache-correct, closed-loop context compression for LLM coding agents. A Rust
> workspace providing a library (`tare-core`), an HTTP proxy (`tare-proxy`) that speaks the Anthropic
> and OpenAI APIs, a CLI (`tare`), an MCP server (`tare-mcp`), Python bindings (PyPI:
> `tare-compress`), and an npm distribution (`tare-ai`).

tare compresses an agent's context window before it reaches the model. It is **lossless by default**
(columnar JSON/log re-encode, exact/envelope dedup, cross-turn IVM delta, supersession, schema-slim,
query-relevance) and **lossy only on opt-in** (row-cap, field-truncate, telegraphic NL, AST code
skeletonization). It is **cache-correct**: only the dynamic suffix after the provider's cache
breakpoint is rewritten, so prefix caches keep hitting. It is **closed-loop**: a per-session
controller adapts aggression from cache-hit-rate (halt), output-verbosity (the compression paradox —
back off), and context-fill (compress harder).

## Usage

- Proxy: `tare-proxy` — env `TARE_UPSTREAM`, `TARE_PORT` (8787), `TARE_CONTEXT_LIMIT` (200000);
  endpoints `/v1/messages` (Anthropic) and `/v1/chat/completions` (OpenAI).
- CLI: `tare compress | slim-schema | compact-lossy | skeletonize | compact-html | compact-csv |
  deref-images | doctor | perf | learn | dashboard | output-savings | update | wrap | unwrap`.
- MCP: `tare-mcp` (stdio) — `tare_compress`, `tare_skeletonize`, `tare_compact_lossy`,
  `tare_deref_images`, reversible `tare_expand`, `tare_stats`, and cross-session memory
  (`tare_remember` / `tare_recall` / `tare_forget` / `tare_memory_stats`).
- Python: `pip install tare-compress` → `import tare` (all core transforms +
  `tare.integrations` framework adapters).
- JS/TS: `npm install tare-ai` → `withTare`, `tareMiddleware`, `startProxy` (+ prebuilt binaries).

## Docs

- [README](README.md): overview, install, benchmarks, comparison, status.
- [CONTRIBUTING](CONTRIBUTING.md), [SECURITY](SECURITY.md), [CHANGELOG](CHANGELOG.md).

## Key facts

- AST code skeletonization: 59–73% token reduction on the committed proof corpus (code reads are the
  dominant sink). Languages: rust, python, js, ts, go, java, c, c++, perl.
- Lossless by default; cache-prefix-boundary aware; bounded body buffering + upstream timeouts.
- MIT licensed. MSRV 1.82.