Terse — the AI agent butler

Cut AI coding agent costs 40–70%. Live-monitor Claude Code, Cursor, Codex and Copilot, stop runaway agents before the next API call, and compress every prompt on-device. macOS & Windows.

🌐 terseai.org  ·  ⬇️ Download  ·  📖 Docs  ·  💸 Token calculator  ·  ⚖️ vs ccusage & others

What is Terse?

Terse (the tool at terseai.org) is an on-device AI agent butler for macOS and Windows, with Chrome and VS Code extensions. It watches the AI coding agents you already run — Claude Code, Cursor, OpenAI Codex, GitHub Copilot CLI, Cline, Windsurf, OpenClaw and Aider — and handles the parts that quietly cost you money:

• Compresses every prompt 40–70% before it hits the API, meaning preserved. 35+ techniques, sub-5ms, code always protected.
• Monitors each agent live — tokens, cost, cache read/write efficiency, burn rate, context fill.
• Stops runaway agents with a budget circuit breaker that pauses (SIGSTOP) or kills (SIGTERM) the process before the next API call.
• Manages your MCP servers — discover, risk-score and toggle without editing JSON.
• Diagnoses waste with ~25 one-click Terse Doctor scans.

Everything runs locally. Your prompts and sessions never leave your machine.

This repository also holds the Terse SDK (MIT) — the token-optimization framework the app is built on — and the benchmark harness behind the 40–70% figure.

How it works

flowchart LR
    U([You prompt an agent]) --> T{{Terse runs on-device}}
    T -->|compress 40-70 percent| S([Smaller prompt])
    T -->|monitor| M([Live tokens, cost, cache, burn])
    T -->|guard| B([Budget breaker stops<br/>runaway agents])
    S --> API([AI API])
    M --> API
    B -.->|before next call| API
    API --> W([Lower bill, same result])

    classDef terse fill:#c6d82c,stroke:#c6d82c,color:#0a0a0a,font-weight:bold;
    classDef node fill:#12140d,stroke:#3a3f26,color:#e8ece0;
    class T terse;
    class U,S,M,B,API,W node;

⭐ If Terse saves you tokens, drop a star — it's the fastest way to help other developers find it.

Quickstart

The app (monitoring, budget breaker, MCP manager, Doctor):

# macOS — download the signed .dmg
open https://github.com/lucaszengool/Terse/releases/latest

Windows build and extensions: see Download.

The SDK (build cost-aware LLM apps yourself — MIT licensed):

git clone https://github.com/Terse-AI/terseai.git
cd terseai
npm run benchmark      # reproduce the 40-70% numbers on your own machine
npm test               # 9 behavioural tests over the public API

import { TerseContext } from './src/index.js';

const ctx = new TerseContext({
  model: 'claude-sonnet-4-6',
  budget: 8000,              // hard token ceiling for the context window
  compression: 'balanced',   // 'soft' | 'balanced' | 'aggressive'
});

const result = await ctx.chat([{ role: 'user', content: 'Explain recursion.' }]);

The @terse-ai/sdk npm package isn't published yet — install from source by cloning this repo.

See it

Overview dashboard Agent health, spend, burn rate & context at a glance.	Token receipts & charts See exactly where your tokens and dollars go.
Doctor — ~25 waste scans Context overflow, duplicate MCP servers, one-click fixes.	Team collaboration Share live agent sessions and hand off work.

Capabilities

Which AI coding agents does Terse monitor?

Eight, auto-detected with no setup:

Claude Code · Cursor · OpenAI Codex · GitHub Copilot CLI · Cline · Windsurf · OpenClaw · Aider

Claude Code goes deepest: exact token counts, cache read/write efficiency, live JSONL streaming, and 30 days of historical backfill. The prompt optimizer works with any AI chat or agent, including ones not on this list.

How is Terse different from ccusage and usage dashboards?

Most tools in this space report what you spent. Terse is built to change it — compression before the call, and a circuit breaker that stops the process before the next one.

→ Full comparison, including when to use something else — ccusage is free, excellent and one command away if all you need is a number.

Download

App: free 30-day trial · $4.99/mo · pricing. SDK: free, MIT.

Why it matters

Three optimization modes

Code blocks, file paths, and technical terms are always protected.

• Soft — typo correction + whitespace only. 100% meaning-safe.
• Normal — removes filler, hedging, politeness padding, meta-language.
• Aggressive — maximum compression: abbreviations, article removal, telegraph style.

Grounded in real research — LLMLingua, Norvig spelling, and selective-context pruning.

Learn more

Guides: What is token optimization · Reduce AI API costs · Claude Code pricing 2026 · AI token pricing comparison · Blog

Compare: Cursor vs Claude Code · Claude Code vs Copilot · Windsurf vs Claude Code · AI coding agent costs

Per-tool: Claude Code · Cursor · ChatGPT · Copilot · Aider · Cline · Windsurf · Codex

In this repo: FAQ · Comparison · SDK reference · Examples · Benchmark · llms.txt

FAQ

Four levers, roughly in order of impact: send fewer tokens (Terse compresses every prompt 40–70% on-device); stop paying for cache misses (Doctor flags cache thrash — reordering a session so the stable prefix stays stable often beats compression); cut MCP tool bloat (unused servers re-send their catalogs on every call); and cap the downside with a hard ceiling, because most large bills are one unwatched loop. More →

Terse auto-detects and monitors 8 agents: Claude Code, Cursor, OpenAI Codex, GitHub Copilot CLI, Cline, Windsurf, OpenClaw, and Aider. Claude Code has the deepest integration (exact token counts, cache efficiency, live JSONL streaming, 30-day history). The prompt optimizer works with any AI chat or agent.

A hard spending limit enforced at the process level rather than reported after the fact. Set a burn-rate, token, or dollar ceiling and Terse escalates from an alert to pausing (SIGSTOP) or killing (SIGTERM) the agent process before the next API call — so a looping agent can't burn hundreds of dollars overnight. Dashboards tell you the money is gone; a circuit breaker keeps it.

No. All compression and analysis run locally in a Rust/JavaScript engine. Your prompts and conversations are never sent to Terse's servers. Optional sign-in only enables subscription and team-sync features.

In Soft and Normal modes, meaning is fully preserved — code blocks, file paths, and technical terms are always protected. Aggressive mode maximizes savings (abbreviations, article removal, telegraph style) for when you want the smallest possible prompt.

40–70% on verbose prompts, less on prompts that were already terse. You can reproduce the numbers yourself: git clone https://github.com/Terse-AI/terseai.git && cd terseai && npm run benchmark. It reports each module separately (text compression, working memory, tool optimization, model routing) rather than one headline number.

The app has a free 30-day trial, then $4.99/month. The Chrome extension has a free tier. The Terse SDK in this repository is MIT licensed and free, benchmark harness included.

Model Context Protocol (MCP) servers add tools to your agent — but bloated or unused tool catalogs quietly add hundreds of tokens to every call, and some servers carry security risk (remote transport, embedded credentials, code execution, unpinned supply chain). Terse discovers every MCP server across your Claude Code / Cursor / Windsurf configs, risk-scores each, and lets you enable or disable them without editing JSON.

The desktop apps ship for macOS and Windows today. The Terse SDK in this repository is plain Node.js and runs anywhere Node 18+ runs, Linux included.

→ Full FAQ

The Terse SDK (MIT)

This repo contains the Terse SDK, a token-optimization framework for building cost-aware LLM apps: context compression, selective/verbatim compressors, working & episodic memory, model routing, and MCP/tool-catalog optimization.

import { linguisticCompress, optimizeTools, ModelRouter } from './src/index.js';

📖 Full SDK reference → SDK.md · examples/ · benchmark/ · Contributing · License

Privacy

All compression and analysis happen on your device (Rust/JS engine). Your prompts and conversations are never sent to Terse's servers. Optional sign-in enables subscription and team-sync only.