PATK MCP Server
Filters verbose terminal output from commands like npm install, pip install, docker build, and pytest, reducing context token consumption for AI agents by condensing logs, removing progress bars, and grouping repeated warnings.
README
PATK — PFC Agent Token Killer MCP Server
Compress noisy terminal output before it fills Claude's context. npm install, docker build, pytest — filtered down to what actually matters.
Why PATK?
Long coding sessions fill Claude's context with terminal noise.
Each npm install dumps 150–200 lines of deprecation warnings and progress bars.
That's ~10,000 tokens of pure noise per session — eating into your context window.
PATK filters that noise before Claude ever sees it.
Real benchmark (5-command session, Haiku API):
| Command | Lines Before | Lines After | Reduction |
|---|---|---|---|
npm install |
52 | 6 | 80% |
pip install -r requirements.txt |
42 | 10 | 77% |
docker build . |
35 | 15 | 57% |
pytest -v |
30 | 21 | 30% |
| Session total | 8,380 tokens | 3,829 tokens | 54% ↓ |
Real Haiku API benchmark. No spin — actual measured values.
What this means:
- Context fills up 2× slower in long sessions
- Claude's summarization is delayed → longer memory of early session details
- At CI/CD scale: meaningful API cost reduction
Works in Sub-Agents Too
Sub-agents (Task tool, Explore agents, parallel agents) have their own isolated context and their own separate API costs. PATK installed once works everywhere:
- Main agent context: protected
- Sub-agent context: also protected (independently)
- n8n pipelines: filter between nodes to halve accumulated data
Add to your Task tool prompts:
"Use patk_safe_execute for all bash commands producing > 15 lines output"
Installation
1. Get your free API key
🔜 API keys not yet live. Watch this repo or open an issue to get notified when pfc-token-killer.com launches — 200 free credits, no payment required.
2. Install dependencies
pip install -r requirements.txt
3. Add to Claude Code MCP config
Claude Code (~/.claude.json or project .mcp.json):
{
"mcpServers": {
"patk": {
"command": "python",
"args": ["/path/to/server.py"],
"env": {
"PATK_API_KEY": "ptk_your_key_here"
}
}
}
}
Or via CLI:
PATK_API_KEY=ptk_your_key_here claude mcp add patk python /path/to/server.py
That's it. No configuration. No extra dependencies beyond requirements.txt.
Tools
| Tool | When to use |
|---|---|
patk_safe_execute |
PRIMARY — Run command + filter in one step. Raw output never enters Claude's context. |
patk_filter_output |
Filter already-captured terminal output (Bash tool fallback) |
patk_check_credits |
Check remaining API credits |
patk_status |
Show configuration and session statistics |
Example: patk_safe_execute
patk_safe_execute(command="npm install", max_lines=50)
→ npm install [✅ Exit 0]
──────────────────────────────────────────────────
added 847 packages in 12.3s
[10× npm warn deprecated — run 'npm audit fix']
// ✂ 80% reduction — 52→6 lines
---
⚡ PATK API | 80% Reduktion | 52 → 6 Zeilen | ~1,200 Tokens gespart
Filter Pipeline
- ANSI removal — Strip colors, cursor codes
- Progress bar detection — Remove
[=====> ] 47%,━━━━━━ 100%, download bars - Pattern condensation — Group similar lines:
[10× npm warn deprecated],[15× pip Collecting/Downloading] - Duplicate compression — Identical lines → single summary
- Timestamp clusters — 20 timestamp log lines → 1 summary
- Entropy scoring — Remove low-information lines; always keep errors/warnings
Compatibility
Works with any MCP-compatible client:
- Claude Code (primary target)
- Cursor
- Cline
- Windsurf
- Any client supporting the MCP protocol
Pricing
| Plan | Credits | Price |
|---|---|---|
| Starter | 200 | Free |
| Hobby | 2,000 | $5 one-time |
| Pro | 10,000 | $15 one-time |
| Enterprise | 50,000 | $49 one-time / seat |
1 credit = 1 filter call. Credits never expire. No subscriptions.
⏳ Paid plans not yet live. Free Starter tier available at launch. Pre-register now for +50% Early Bird bonus credits on paid plans.
Get your key
→ Register at pfc-token-killer.com (coming soon) → Or watch this repo — we'll announce here when keys go live
Questions / Support
→ Open an Issue — fastest way to reach us
Built by ForgeBuddy & Dante · 2026
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。