Skip to content

getting-started.md

Latest commit

 

History

History
69 lines (51 loc) · 2.42 KB

getting-started.md

File metadata and controls

69 lines (51 loc) · 2.42 KB
Raw

Getting started

Build

git clone https://github.com/mstuart/tare && cd tare
cargo build --release          # builds target/release/{tare, tare-proxy}

The default build uses keyword/symbol-based relevance and has no external model dependency. To enable neural semantic relevance (exact cosine ranking via fastembed — downloads an embedding model on first use):

cargo build --release --features neural-embed

Or with Docker:

docker compose up --build      # runs tare-proxy on :8787

As a proxy

Point your agent's API base URL at tare; it forwards to the real upstream and compresses on the way.

TARE_UPSTREAM=https://api.anthropic.com TARE_PORT=8787 ./target/release/tare-proxy

It speaks Anthropic (/v1/messages) and OpenAI (/v1/chat/completions).

Env var Default Meaning
TARE_UPSTREAM https://api.anthropic.com upstream API base URL
TARE_PORT 8787 listen port
TARE_RECENCY 4 tool outputs always kept regardless of relevance
TARE_ENABLED true set 0/false for byte-exact passthrough
TARE_CONTEXT_LIMIT 200000 model context window (drives the fill-based aggression dial)
TARE_OUTPUT_HOLDOUT 0 fraction of sessions that bypass compression (A/B baseline for tare output-savings)
TARE_LOG unset set it to log one line per turn with the compression report

Response headers report what it did: x-tare-input-tokens, x-tare-net-tokens, x-tare-dropped, x-tare-aggression, x-tare-verbosity-spike, x-tare-halted.

As a CLI

cat ctx.json   | tare compress --task "fix the auth bug"
cat big.rs     | tare skeletonize --path big.rs
ps aux         | tare compact-lossy --max-rows 30 --max-field 110

See the CLI reference for all subcommands.

As an MCP server

tare-mcp is a stdio MCP server. Point any MCP client at the tare-mcp binary; it exposes tare_skeletonize, tare_compact_lossy, tare_compress, tare_deref_images, tare_stats, and a reversible tare_expand — when a tool compacts something it returns an id, and tare_expand({id}) returns the exact original (CCR-style retrieval), so the agent can drill back in on demand — plus persistent cross-session memory: tare_remember, tare_recall, tare_forget, tare_memory_stats.

// example MCP client config
{ "mcpServers": { "tare": { "command": "tare-mcp" } } }