Cut 40-70% of your AI token costs with on-device optimization — then give your whole team full visibility with Terse Cloud. Token analytics by developer, project, and tool.
First Launch — macOS Security Step (one-time only)
macOS blocks unsigned apps by default. Pick whichever method works for you:
xattr -cr /Applications/Terse.app && /Applications/Terse.app/Contents/MacOS/terse 2>/dev/null &
Terse is a macOS & iOS token optimization tool that reduces AI prompt costs by 40-70% through 35+ intelligent compression techniques — including git diff compression, prompt caching awareness, conversation history summarization, model-specific optimization, duplicate tool call detection, redundant file read flagging, and real-time per-turn cost tracking across Claude Code, Cursor, ChatGPT, OpenClaw, and Aider. Built on research from LLMLingua (EMNLP 2023), Norvig spelling correction, and selective context pruning.
Updated March 2026 · v1.2.0
Intercepts prompts and agent commands before they hit the API — strips filler, fixes typos, shortens phrases, compresses verbose text. 35+ techniques including output compression: log deduplication, stack trace collapsing, JSON schema extraction, and terminal noise stripping. 40-80% fewer input tokens.
Auto-detects Claude Code, OpenClaw, Cursor, and Aider. Tails JSONL session logs in real time — tracking input/output tokens, cache efficiency, tool call overhead, duplicate calls, redundant file reads, and per-turn cost with model-aware pricing.
Detects duplicate tool calls, flags files read multiple times, estimates compressible tool results (git 85%, tests 90%, builds 80%), and catches typos that cause costly retry loops. See how to cut AI API costs by 60-80%.
Detects git diff output and compresses to only changed lines + 1 line of context. Typical 70% reduction — critical for Claude Code and Cursor workflows where diffs dominate tool output.
Detects stable prefixes and repeated content blocks that benefit from prompt caching — the #1 cost reduction technique in 2026. Saves 50-90% on repeated context across agent turns.
Auto-summarizes old turns in long agent sessions: "User: A B C, Assistant: X Y Z" → "Earlier: Discussed A, concluded X." Massive savings in multi-turn workflows — up to 80% on history context.
Claude handles implicit reasoning — Terse compresses more aggressively. GPT-4 needs explicit instructions — Terse preserves structure. Automatic model detection adapts compression strategy per provider.
A single agent task can consume 50x more tokens than a chat message. Terse runs 5 parallel optimization strategies: compressing your prompts, detecting duplicate tool calls, flagging redundant file reads, estimating compressible tool results, and auto-generating CLAUDE.md rules that teach agents to waste fewer tokens next session.
No config files. No API keys to paste. No terminal commands. Open Terse, see your running agent, click Connect — optimization starts immediately. Every message compressed, every session tracked, every wasted token flagged.
RTK (Reduce Token Kit) and similar libraries offer basic prompt trimming. Terse goes far beyond with agent-aware optimization, real-time monitoring, and output compression — techniques RTK doesn't support.
RTK = Reduce Token Kit. Comparison based on publicly documented features as of March 2026.
Connect Terse to Chrome, Cursor, VS Code, OpenClaw, or any terminal — it auto-detects text fields via macOS Accessibility and agent sessions via process scanning. No plugins to install. Agent sessions are detected automatically every 5 seconds.
As you type, Terse rewrites your prompt in real time — fixing typos, stripping filler, compressing verbose phrasing. But it doesn't stop there: when agents take actions, Terse optimizes their commands too. Every token saved — from your prompts and agent operations — means lower cost and better responses.
Filler removal, question-to-imperative, Jaccard deduplication, telegraph compression — each technique targets a different source of token waste. Applied to your prompts and agent commands alike.
Different contexts need different levels. Soft for careful prompts, Normal for everyday chat, Aggressive for agent sessions where a single task can burn through thousands of tokens.
Terse auto-detects Claude Code, OpenClaw, Aider, and Cursor Agent via process scanning every 5 seconds. It tails JSONL session logs in real time with model-aware pricing (Opus, Sonnet, Haiku) — giving you full visibility into where every token goes.
Terse runs a local proxy on port 7860 that intercepts every API call. Simple tasks — short prompts, lookups, edits — are silently rerouted from Opus ($15/MTok) to Sonnet ($3/MTok). Complex tasks stay on Opus. You pay 80% less with zero code changes.
Tested on real ChatGPT prompts, Claude Code agent sessions, and multi-turn agent workflows. Clean technical prompts pass untouched. Verbose prompts and agent messages see 40-70% reduction. Combined with tool overhead savings, total session reduction reaches 30-60%.
Side-by-side comparison on actual prompts and agent commands.
Even at $200/mo, every tool has caps. Terse compresses prompts so your limits stretch further — and Terse Cloud gives your team full visibility into exactly where tokens are going.
Without Terse, CLI output and verbose prompts alone can overflow a 200K context window. Based on avg 3,500 tokens/command measured across real coding sessions.
No matter which AI tool your team uses, token costs add up fast. Terse compresses what goes in — and Terse Cloud gives you team-wide analytics to track spend by developer, project, and tool.
Pricing verified Feb 2026. Limits vary by usage and plan. Terse savings based on avg 89% compression across real prompts and CLI outputs.
Pick a pal that celebrates every token you save — it reacts to each tool call, eats your savings, and keeps you company through long coding sessions.
20 pals available · Click to poke · Unlock new ones as you save tokens
Engineers and AI power users cutting costs and gaining visibility into their token usage.
Grounded in LLMLingua, Norvig spelling, selective context pruning, and real-world agent session analysis across thousands of Claude Code turns.
Integrate token optimization directly into your vibe coding project. One API call — 30–60% fewer tokens, same meaning, same results.
Sign in and generate a tsk_... developer key from your dashboard. Free tier included.
POST your prompt to /api/v1/optimize — get back a trimmed version with token count saved.
Use /api/v1/scan to find every LLM call site in your project and see where tokens are wasted.
Free tier included with every account. No credit card for free. Upgrade for higher limits and batch access.
No credit card needed.
Cancel anytime. No trial.
Teams & high-volume use.
Built with Terse API? Publish your project to get traffic, users, and visibility in the vibe coding developer community.
Orchestrates Claude agents for complex tasks. Terse API cuts context window usage by 41% per turn, enabling longer sessions without hitting limits.
A library of 200+ production-ready prompts. Uses Terse API to auto-compress each prompt at runtime, reducing developer API bills by $80/mo on average.
Batteries-included Python scaffold for building LLM agents. Terse integration is built-in — every outbound message is auto-optimized before the API call.
Have a project using Terse API?
Visibility and control over your team's AI coding costs. See who's using what, how much is saved, and where tokens are wasted — across every tool.
Dashboard by developer, project, and tool — Mac, Windows, Chrome, VS Code, iOS.
"Your team saved $4,200 this month" — with per-developer and per-project breakdowns.
Smart notifications when a developer is approaching heavy usage thresholds.
Distribute API tokens, add or remove developers, and audit all activity in one place.
Free for open source. Teams from $15 / developer / month.
Every plan includes a 30-day free trial. No charge until your trial ends.
For developers running agent sessions daily. Unlimited prompts, multi-session monitoring.
For teams and power users. Unlimited everything, priority support.
Everything you need to know about token optimization and how Terse saves you money.
Research, guides, and analysis on reducing LLM token costs — from prompt caching to model routing to agent session efficiency.
Prompt caching is now the single highest-ROI technique for reducing AI agent costs. Cached tokens cost 10× less than standard input tokens — here's how to structure every session for maximum cache hit rate.
When Claude Code runs git diff, the output can consume 30–60K tokens per turn. Compressing to changed lines + 1 line of context is the fastest context win in any agent workflow.
Analysis of 500 real Claude Code sessions found 34% of Read tool calls were redundant. At $15/MTok for Opus, each repeated file read wastes ~$0.006 — small per turn, painful at scale.
Not every agent task needs Claude Opus. Single-file edits, test generation, and short lookups perform identically on Sonnet — saving teams $200+/month with zero quality loss on routine tasks.
LLMLingua uses perplexity scoring to drop tokens at inference time. Deterministic rule pipelines work at the prompt layer before the API call. We tested both on 1,000 real developer prompts.
A single typo in an agent prompt can trigger a clarification turn, adding 200–500 tokens of overhead. Typo-caused retries account for 8–15% of total token spend across a typical week of Claude Code usage.
Optimize every prompt. Monitor every agent session. And give your whole team visibility with Terse Cloud — analytics by developer, project, and tool.
First Launch — macOS Security Step (one-time only)
macOS blocks unsigned apps by default. Pick whichever method works for you:
xattr -cr /Applications/Terse.app && /Applications/Terse.app/Contents/MacOS/terse 2>/dev/null &