tare

Cut your coding agent's context 44–73% — lossless by default, cache-safe, output-aware.

Install · Proof · Compared to · Claude / MCP · Integrations · Status

tare sits between your agent and the model API and shrinks the context window. Three properties make
it different:

• Lossless by default — tool output, logs, and JSON are re-encoded into a denser equivalent
  form. Information is dropped only when you opt in (row-caps, telegraphic NL, AST code
  skeletonization).
• Cache-correct — provider prefix caches discount cached tokens ~10×, and one rewritten byte in
  the prefix forfeits it. tare detects the cache breakpoint and compresses only the dynamic suffix.
• Output-aware — over-compression makes models compensate with verbose output, so total cost can
  rise as input falls. tare watches output tokens per turn and backs off when they spike.

What you get

• tare-proxy — point your agent's base URL at it; zero code changes. Speaks Anthropic
  (/v1/messages) and OpenAI (/v1/chat/completions).
• tare CLI — compress, skeletonize, compact-lossy, and 12 more subcommands
  (full reference).
• tare-mcp — compression + persistent cross-session memory as MCP tools, for any MCP client.
• Libraries — tare-core (Rust), tare-compress (Python),
  tare-ai (JS/TS).

Proof

All numbers are measured by running target/release/tare on the committed corpus with tiktoken
o200k_base (v0.13.0). Results live in crates/tare-bench/results/; reproduce with
python3 crates/tare-bench/run_proof.py.

Code reads are ~67–76% of a coding agent's tokens (SWE-Pruner, ACL 2026),
so skeletonization is the biggest lever. Head-to-head harnesses against Headroom, LLMLingua-2,
lean-ctx, and RTK live in crates/tare-bench/benchmarks/.

Compared to

¹ Output-aware = reads the model's output token count each turn and reduces compression aggression
when verbosity spikes (the x-tare-verbosity-spike signal in
crates/tare-proxy/src/server.rs).

When to use · When to skip

Great fit if you…

• run coding agents and want savings without losing information by default
• care about the provider cache staying warm (tare won't break the prefix)
• want code reads compressed structurally (signatures kept, bodies elidable on demand)

Skip it if you…

• only use a single provider's native compaction and don't need a cross-provider proxy
• run in a sandbox where a local proxy process can't run

Get started (60 seconds)

# 1 — install (no Rust toolchain needed)
curl -fsSL https://raw.githubusercontent.com/mstuart/tare/main/install.sh | sh   # → ~/.local/bin
# or:  npm install -g tare-ai
# or:  docker pull ghcr.io/mstuart/tare
# or:  cargo install tare-cli
# or:  git clone https://github.com/mstuart/tare && cd tare && cargo build --release

# 2 — run as a proxy (point your agent's base URL at http://localhost:8787)
TARE_UPSTREAM=https://api.anthropic.com tare-proxy

# 3 — or use the CLI on any stdin
cat big.rs | tare skeletonize --path big.rs    # drop fn bodies, keep structure
ps aux     | tare compact-lossy --max-rows 30 --max-field 110

Response headers (x-tare-input-tokens, x-tare-net-tokens, x-tare-dropped, x-tare-aggression,
x-tare-verbosity-spike, x-tare-halted) report what each turn did; GET /admin/stats and
POST /admin/runtime-env expose live stats and hot config when TARE_ADMIN_TOKEN is set; clients
must send the token in x-tare-admin-token. Details: getting started.

Integrations

# pip install tare-compress
import tare
out = tare.compress(blocks_json, task="fix the login bug")   # in-process, no proxy needed

// npm install tare-ai
import { withTare } from "tare-ai";
const client = new Anthropic(withTare({ apiKey: "..." }));   // routes through the local proxy

Python ships all core transforms in-process plus adapters for litellm, ASGI, LangChain, Agno, and
Strands; JS/TS ships proxy helpers for the Anthropic/OpenAI SDKs and the Vercel AI SDK.
Full adapter reference: docs/integrations.md.

Use with your Claude subscription — MCP, no API key

The proxy forwards whatever auth the client sends — a billable x-api-key, or your Claude Pro/Max
subscription OAuth token when you point Claude Code's ANTHROPIC_BASE_URL at it
(scripts/live-smoke-sub.sh runs exactly this round-trip). Prefer no base-URL change at all? Run
tare-mcp: a local stdio server your agent calls as tools — it never calls the model itself, so it
needs no API key.

# easiest — no build; npx fetches the prebuilt binary on first run:
claude mcp add tare -s user -- npx -y -p tare-ai tare-mcp

{
  "mcpServers": {
    "tare": { "command": "npx", "args": ["-y", "-p", "tare-ai", "tare-mcp"] }
  }
}

Paste the same block into any MCP client (Cursor, Codex, Claude Desktop's
claude_desktop_config.json, …). It exposes 10 tools: compression (tare_compress,
tare_skeletonize, tare_compact_lossy, tare_deref_images), a reversible tare_expand,
tare_stats, and cross-session memory — full list.

Wrap your agent

One command starts the proxy and launches your CLI agent through it — ENV-based and ephemeral:

tare wrap claude               # start proxy + launch Claude Code through it
tare wrap codex --port 9000    # 12 agents supported: claude, codex, aider, goose, …
tare wrap claude --print       # dry-run: show what would run

Full agent matrix and modes: docs/cli.md.

Status & limitations

• Tested — verified end-to-end against the live Anthropic API (a full proxy round-trip on a
  Claude subscription, plus the MCP server over real stdio JSON-RPC), on top of 228 unit,
  integration, and property tests; fmt / clippy -D warnings / cargo deny gate every commit.
• Deploy it as a local sidecar — tare runs next to your agent and forwards your credentials
  upstream without logging or persisting them; treat it as a trusted component on your own machine,
  not shared multi-tenant infrastructure (SECURITY.md). Startup failures exit with a
  clear [tare-proxy] fatal: … message, never a panic backtrace.

Known edges

• Proxy and CLI token counts are approximate (tare-tokenize, chars/4); the benchmark numbers above
  are measured with tiktoken.
• The context-fill signal counts the serialized request (incl. JSON envelope), so it slightly
  over-estimates fill — conservative, errs toward compressing sooner.
• A >2 MB streaming response whose final usage event straddles the 64 KB tail buffer may skip
  one verbosity sample (non-fatal).

Non-goals — no trained ML text-compressor (no weights to download, no inference latency) and no
audio: transcribe externally and feed the transcript through tare compress.

Architecture

Nine crates: engine (tare-core), proxy + closed-loop controller (tare-proxy), CLI, MCP server,
SQLite memory, tokenizer, cache models, Python bindings, bench harness. Optional neural-embed
feature swaps keyword relevance for exact-cosine neural embeddings.
Diagram and crate guide: docs/architecture.md.

Contributing

See CONTRIBUTING.md. CI runs cargo fmt --check, cargo clippy -D warnings,
cargo test, and a release build — please make sure those pass locally.

License

MIT — see LICENSE.