🚀 Agent Boost Kit

💡 Why Agent Boost Kit?

Coding agents often fail because they skip steps, lose context, or generate too much "filler" text. agent-boost-kit solves this by giving your agent rails, ground truth context, and token-saving superpowers.

• No LLM API keys required. It works inside the environment of the agent you already use.
• Works with everything. Aider, Claude Code, Cursor, Cline, VS Code, and more.
• Pay less, get more. Spend 80% less on tokens while maintaining 90%+ of heavy-model quality.

🚀 Quick Start

1. Install

npm install -g agent-boost-kit
# or run directly
npx agent-boost-kit init

🔄 Modes — Choose Your Setup

agent-boost-kit comes with three distinct modes designed for different AI model tiers and budgets. During init, you can choose a mode or switch later.

1. 🚀 Boost Mode (Focus: Code Quality)

Best for: Free or weak models (e.g., Gemini 2.0 Flash, GPT-4o-mini, local 7B models).
Goal: Make cheap models code like Senior Developers.

• ✅ Flash Enhancer: Auto-injects expert roles, structure protocols, and forces retries on bad answers.
• ✅ Strict Verifier: Blocks the agent from finishing until Lint, Tests, and Build pass.
• ✅ Codemap & Lessons: Gives the agent deep repo context and remembers past mistakes.
• ❌ Token saving features are disabled (quality takes priority over API cost).

2. 💰 Saver Mode (Focus: Cost Reduction)

Best for: Premium, expensive models (e.g., Claude 3.5 Sonnet, GPT-4o).
Goal: Cut API costs by up to 80% without losing context.

• ✅ CLI Proxy: Compresses noisy terminal output (tests, git diffs) by 90% before the AI reads it.
• ✅ Input Compression: Removes filler words and trims history context.
• ✅ Caveman Output: Forces the LLM to write zero fluff (only code, no conversational filler).
• ✅ Smart Router: Auto-routes simple tasks to free models and saves premium quota for complex tasks.
• ✅ Auto-Hook Injection: install command auto-detects your AI tool and injects mandatory compression rules.
• ✅ Project Intelligence: anatomy.md indexes all files with token counts; cerebrum.md tracks mistakes to avoid repeating them.
• ❌ Quality boosting features are disabled (Premium models usually don't need them).

3. ⚡ Both Mode (Default)

Best for: Any model.
Goal: The best of both worlds.

• ✅ All Quality Enforcement features are ON.
• ✅ All Token Saving features are ON.

Managing Modes

# Set a specific mode
npx agent-boost-kit mode set boost    # weak model + quality only
npx agent-boost-kit mode set saver    # strong model + token saving (auto-installs hooks!)
npx agent-boost-kit mode set both     # all features ON (default after init)

# Check what's active
npx agent-boost-kit mode status

Note: both is selected by default when running init (just press Enter). You can switch anytime without breaking your project.

Saver/Both mode auto-setup: When you switch to saver or both, agent-boost-kit automatically scans your project (builds anatomy.md file index) and injects compression hooks into your AI tools. No extra commands needed.

💎 Key Features

| Feature | Description | | :--- | :--- | | 🧠 Flash Enhancer | Automatically detects the task domain (14 supported) and task type (debug, refactor, feature) to inject expert-role prompts. Includes an evaluateAndRetry feedback loop that penalizes bad AI answers and forces retries. | | 📉 Token Saver | 5-layer compression system. Includes a Smart CLI Proxy that detects 25+ commands to compress test outputs and git diffs, saving ~90% terminal noise. Auto-hooks into AI tools via install command. | | 🚦 Model Router | Cost-aware routing based on task complexity. Simple tasks go to free models, complex architecture tasks automatically escalate to Premium (Claude/GPT-4). Includes a live terminal dashboard. | | 🗺️ Auto-Codemap | Builds a deep AST-based map of your repo so agents don't have to guess. | | 📜 Lessons (RAG) | Your agent learns from its mistakes. Every fix becomes a permanent rule for future tasks. | | 🎯 Blast Radius | Analyzes the impact of code changes before they happen using Git history and imports. | | 🤖 Auto Self-Review | Generates a structured self-review prompt by analyzing diffs and codemaps, forcing agents to check for bugs and edge cases before declaring a task done. | | 🛡️ Verifier | A strict "Exit 0" contract. The agent isn't done until Lint + Types + Tests pass. | | ✨ Zero-Token Auto-Fix | Automatically detects and runs --fix for basic syntax errors locally before returning failures to the LLM. | | 📂 Project Intelligence | anatomy.md indexes every file with token counts so the AI never re-reads files. cerebrum.md tracks known mistakes to prevent repeats. |

🤝 Supported Agents

| Adapter | Target Tool | What Gets Written | | :--- | :--- | :--- | | antigravity | Antigravity IDE | GEMINI.md (always-on, mode-aware) + .agent-boost/skills/ referenced inside | | opencode | OpenCode | AGENT.md + opencode.json ($schema, instructions, rules, default model) | | claude-code | Claude CLI | CLAUDE.md + .claude/skills/ + .claude/commands/ + settings.json hook | | cursor | Cursor IDE | .cursor/rules/agent-boost-kit.mdc + Verifier | | cline | Cline / Roo Code | .clinerules + MCP Integration | | aider | Aider CLI | .aider.conf.yml + Read-logs | | vscode | VS Code / Copilot | .vscode/tasks.json + AGENT.md | | generic | Any tool | AGENT.md + .agent-boost/skills/ + .agent-boost/commands/ |

🛠️ Commands Reference

# Setup
npx agent-boost-kit init              # Scaffold + choose mode
npx agent-boost-kit mode set saver    # Switch to saver (auto-runs install + scan)
npx agent-boost-kit mode set boost    # Switch to quality mode (weak models)
npx agent-boost-kit mode status       # Show active mode + feature table

# Token Saving (Saver Mode)
npx agent-boost-kit compress "prompt" # Compress a prompt & score complexity
npx agent-boost-kit proxy "npm test"  # Wrap a command to save output tokens (25+ tools supported)
npx agent-boost-kit token-status      # View live savings & model status
npx agent-boost-kit dashboard         # Open web dashboard for savings

# Quality (Boost Mode)
npx agent-boost-kit codemap           # Rebuild project index
npx agent-boost-kit verify            # Run Lint + Tests + Build
npx agent-boost-kit self-review       # Generate auto-review prompt for current diff

# Manual (optional — auto-runs on mode set saver/both)
npx agent-boost-kit install           # Re-inject hooks into AI tools
npx agent-boost-kit scan              # Re-scan project file index

# Cleanup
npx agent-boost-kit uninstall         # Remove all agent-boost-kit files

⚖️ License

Distributed under the MIT License. See LICENSE for more information.