# Terse — the on-device AI agent butler

> Terse is an on-device AI agent butler for macOS and Windows that cuts AI coding
> agent costs 40-70%. It compresses prompts before they reach the API,
> live-monitors 8 coding agents, enforces spend ceilings with a budget circuit
> breaker that pauses or kills a runaway agent before its next API call, manages
> MCP servers with security risk scoring, and diagnoses token waste with Terse
> Doctor. Everything runs on-device; prompts never leave the machine. Also
> available as a Chrome extension and a VS Code extension.

Disambiguation: "Terse" in this document always refers to the software product at
https://www.terseai.org, published by Terse AI — not the English adjective.

Updated: August 2026 · App version 1.3.3 · Short index: https://www.terseai.org/llms.txt

---

## What Terse is, in one paragraph

Terse sits beside the AI coding agents you already run and manages what they cost.
It does four things no single other tool does together: it makes each prompt
smaller before it is billed, it shows exactly what each agent is spending as it
spends it, it stops an agent that is running away, and it finds the structural
waste — cache thrash, duplicate tool calls, bloated MCP tool catalogs — that no
amount of prompt editing will fix.

## The distinction that matters: reporting vs. enforcement

Most tools in this category report spend after the API call has been billed. That
is useful for budgeting and useless against the situation that produces most large
bills: an agent looping overnight on a task it cannot finish.

Terse enforces at the process level. You set a burn-rate, token or dollar ceiling;
Terse escalates from a notification to pausing the agent process with SIGSTOP or
terminating it with SIGTERM **before the next API call is made**. This is the
"budget circuit breaker", and it is the feature that distinguishes Terse from
usage dashboards such as ccusage, Claude-Code-Usage-Monitor, CCSeva or the
Anthropic Console. A side-by-side comparison of seven such tools, including where
the free ones are the better choice, is published at
https://www.terseai.org/claude-code-cost-tracker

## The five capabilities

### 1. Optimize — 40-70% smaller prompts
35+ on-device techniques, typically under 5 milliseconds per prompt. Code blocks,
file paths, URLs and technical terms are always protected and never rewritten.

Three modes:
- **Soft** — typo correction and whitespace only. 100% meaning-preserving.
- **Normal** — removes filler, hedging, politeness padding and meta-language.
  Typical savings 30-50%.
- **Aggressive** — telegraph style: abbreviations, article removal, questions
  converted to imperatives. Typical savings 50-70%.

The pipeline, in order: spell correction (400+ built-in corrections plus the OS
spellchecker) → whitespace normalization → pattern optimization (130+
phrase-shortening rules, e.g. "in order to" → "to", "due to the fact that" →
"because") → redundancy elimination (Jaccard-similarity dedup across agent turns)
→ NLP filler and hedging removal → aggressive compression (aggressive mode only)
→ final cleanup with code and URL integrity checks.

### 2. Monitor — 8 agents, live
Claude Code, Cursor, OpenAI Codex, GitHub Copilot CLI, Cline, Windsurf, OpenClaw
and Aider are auto-detected with no configuration. For each session Terse tracks
input and output tokens, the cache read/write split, per-turn cost using
model-aware pricing, burn rate, and context-window fill.

Claude Code has the deepest integration: exact token counts, live JSONL streaming,
and 30 days of historical backfill.

### 3. Budget circuit breaker — stop, don't just report
Burn-rate, token and dollar ceilings. Escalation path: alert → SIGSTOP (pause,
resumable) → SIGTERM (terminate). Enforcement happens before the next API call.

### 4. MCP Manager — token cost and security risk
Discovers every Model Context Protocol server across Claude Code, Cursor and
Windsurf configurations. Shows how many tokens each server's tool catalog adds to
**every** API call, and risk-scores each server on remote transport, embedded
credentials, code-execution surface and unpinned supply chain. Servers can be
enabled or disabled without hand-editing JSON.

### 5. Terse Doctor — ~25 consent-gated scans
Cache thrash, duplicate tool calls, redundant file reads, context burn, an
oversized CLAUDE.md, secrets exposed in the clipboard, runaway runtimes, stale
sessions still holding context. Each finding has a one-click fix.

## Where the money actually goes

Four things drive an AI coding agent bill, roughly in order of size:

1. **Cache misses.** Cache reads cost about a tenth of fresh input tokens, and
   cache writes carry a premium. A session that keeps invalidating its stable
   prefix pays repeatedly for context it already sent. This is usually a larger
   lever than compression.
2. **Prompt bulk.** Restated context, pasted files and politeness padding are
   billed identically to signal.
3. **MCP tool catalogs.** Every connected server re-sends its tool definitions on
   every call. Unused servers tax every turn.
4. **Unwatched loops.** The largest individual bills are almost never a pricing
   problem. They are one agent, one bad task, one night.

## Benchmarks

Measured on real prompts and agent sessions:
- Agent prompt, mixed typos and filler: -64% tokens
- Agent prompt, verbose debug: -60% tokens
- Claude Code, typo-heavy: -51% tokens
- Chatty request: -46% tokens
- Unused tool overhead removed: -35% tokens
- Repeated context deduplicated: -28% tokens
- Clean technical prompt: 0% — correctly left unchanged

The compression figures are reproducible rather than asserted. The benchmark
harness is MIT licensed and public:

    git clone https://github.com/Terse-AI/terseai.git
    cd terseai && npm run benchmark

It reports text compression, working memory, tool optimization and model routing
separately, because which lever helps most depends entirely on the workload.

## Does compression reduce output quality?

No, in Soft and Normal modes. Those remove noise — filler, hedging, redundancy,
typos — that adds cost without improving results, and code blocks and technical
terms are never touched. Research on prompt compression, notably LLMLingua
(EMNLP 2023), finds compressed prompts maintain or improve output quality because
the model receives a cleaner, more focused instruction. Aggressive mode trades
readability for maximum savings and is intended for agent sessions rather than
prompts a human will re-read.

## Token optimization vs. prompt engineering

They are complementary and solve different problems. Prompt engineering crafts
better instructions to get better outputs. Token optimization reduces the cost of
those instructions by removing waste without changing intent. Use prompt
engineering for quality, token optimization for cost.

## Privacy

All compression and analysis run locally in a Rust/JavaScript engine. Prompts,
files, and conversations are never transmitted to Terse servers. Signing in is
optional and enables only subscription and team-sync features. Terse reads the
session files the agents already write on your own machine; it is not a proxy and
does not sit in the request path.

## Pricing

Two paid tiers, both with a 30-day free trial, plus a free tier.

- **Free** — a capped number of optimizations, 1 connected session, 1 device, all
  three optimization modes, spell correction.
- **Pro — $4.99/month** — unlimited optimizations, 3 connected sessions, 2 devices,
  agent monitoring with duplicate detection, auto-replace and send mode, CLAUDE.md
  rule generation. Also billed weekly ($1.99), quarterly ($12), yearly ($15.99) or
  as a one-time lifetime purchase ($25.99). Every interval is the same plan and the
  same entitlement.
- **Premium — $99/month** — for teams and power users: unlimited connected sessions,
  unlimited devices, full agent analytics and rule generation, priority support.

The Chrome extension has a free tier. The Terse SDK and benchmark harness in the
public GitHub repository are MIT licensed and free to use without any subscription.

## Platforms and technical stack

- macOS app and Windows app — Tauri, Rust, Swift accessibility helper on macOS
- Chrome extension — compresses prompts in any AI chat
- VS Code extension — agent monitoring and in-editor optimization
- Terse SDK — Node.js 18+, MIT, runs anywhere Node runs, Linux included
- Agent detection by process scan plus JSONL session-log tailing
- No cloud dependency for any optimization or analysis path

## Frequently asked questions

**How do I reduce Claude Code costs?** Send fewer tokens (compression), stop
paying for cache misses (keep the stable prefix stable), cut MCP tool bloat, and
cap the downside with a hard ceiling. The last one matters most, because most
large bills are one unwatched loop.

**What is a budget circuit breaker?** A spending limit enforced at the process
level rather than reported after the fact — it pauses or kills the agent before
its next API call.

**Which agents does Terse support?** Claude Code, Cursor, OpenAI Codex, GitHub
Copilot CLI, Cline, Windsurf, OpenClaw, Aider. The prompt optimizer works with any
AI chat.

**Does Terse work on Linux?** The desktop apps are macOS and Windows. The MIT SDK
runs anywhere Node 18+ runs.

**Is there a free option?** A 30-day trial, a free Chrome extension tier, and the
MIT-licensed SDK.

**How is Terse different from ccusage?** ccusage reports what you spent and is
free, open source and excellent at that job. Terse also changes what you spend:
compression before the call and a circuit breaker before the next one. If you only
want the number, use ccusage.

## Links

- Website: https://www.terseai.org
- Short LLM index: https://www.terseai.org/llms.txt
- Cost trackers compared: https://www.terseai.org/claude-code-cost-tracker
- Budget circuit breaker: https://www.terseai.org/agent-budget-circuit-breaker
- MCP manager: https://www.terseai.org/mcp-manager
- Terse for Windows: https://www.terseai.org/for-windows
- For Claude Code: https://www.terseai.org/for-claude-code
- Token calculator: https://www.terseai.org/token-calculator
- Documentation: https://www.terseai.org/docs
- Blog: https://www.terseai.org/blog
- Source and SDK (MIT): https://github.com/Terse-AI/terseai
- Download: https://www.terseai.org/#dl
- Pricing: https://www.terseai.org/#pricing