Every message you send to Claude includes the full conversation history. As sessions grow, you're paying for the same tokens over and over — assistant affirmations, verbose phrasing, structural filler that the model doesn't need to do its job.

Terse strips it before it reaches the API. Transparently, in microseconds, with no code changes.

"Certainly! I'd be happy to help. In order to fix this issue,
 you should take into consideration the edge cases. I hope this helps!"

→  "Fix this issue, consider the edge cases."

   30 tokens → 12 tokens  (-60%)

curl (macOS and Linux):

curl -fsSL https://raw.githubusercontent.com/fwdai/terse/main/install.sh | bash

cargo:

cargo install terse

Manual: download a pre-built binary from Releases.

Terse runs as a local proxy between your CLI and the Anthropic API. Claude Code (and any Anthropic SDK client) respects ANTHROPIC_BASE_URL — set it to http://localhost:3847 and every request is automatically compressed before forwarding. The current message is always sent verbatim; only history is compressed.

Claude Code  →  terse proxy (compress history)  →  api.anthropic.com

One-time setup:

terse install       # hooks into your shell — proxy auto-starts when you run claude
source ~/.zshrc
claude              # that's it, compression is live

No API key changes. No wrapper scripts. Just run claude as usual.

Three tiers, applied in order:

Code blocks, inline code, and URLs are never touched.

Typical savings:

terse gains          # total tokens saved across all sessions
terse gains --watch  # live dashboard, updates as you work

══════════════════════════════════════════════════════
  TOTAL SAVINGS (12 sessions)
══════════════════════════════════════════════════════

  Total calls:      147
  Tokens processed: 2.4M
  Tokens saved:     890.3K (37.1%)

  Efficiency  [████████████████████░░░░░░░░ ] 37.1%

terse proxy status   # see active sessions and per-session savings
terse proxy stop     # stop the proxy
terse upgrade        # update to the latest release

Auto-created at ~/.terse/config.json on first run:

{
  "mode": "trim",
  "tokenizer": "tiktoken",
  "proxy": { "port": 3847 }
}

mode: trim (lossless) · compress (light NLP) · rewrite (planned)

• English only. Other languages pass through unmodified.
• Current message is never compressed. Only history. Your prompt reaches the model exactly as written.
• No external calls. Compression is local, deterministic, sub-millisecond.

Both implementations are verified against the same test fixtures.