PETP

MCP Tool Server Β· Task Orchestration Β· LLM-Native Β· AI-Powered Generation

Turn any workflow into an MCP tool β€” callable by any Agent.

PETP is a pipeline execution runtime that exposes every automation as a typed MCP tool. AI agents can discover, invoke, and chain your real business operations β€” while processors inside each task can call LLMs themselves for intelligent decision-making. Now with built-in OCR, captcha handling, and conditional flow control.

Explore Architecture Browse Shared Files
LLM Inside processors can call LLMs
Any Scenario personal & enterprise toolsets
How It Works
1

Define

Configure tasks visually in GUI or write YAML directly

2

Run

Desktop, headless service, or Docker container

3

Integrate

Expose as MCP tools for AI agents to discover and invoke

Core Capabilities

Flexible Orchestration

Chain any mix of processors in a single execution. Conditional jump with GO_TO_TASK, loop with programmatic break / continue via loop_condition, or branch on LLM output β€” all in YAML, no code required.

LLM Inside Tasks

Any task step can invoke an LLM processor: summarise, classify, generate, or decide. Results flow as variables into downstream steps.

Personal & Enterprise

Run a personal assistant on your laptop, or deploy enterprise tool-sets on Docker. Same engine, same YAML, different scale.

Visual Editor with Undo / Redo

GUI editor with full undo / redo and snapshot history. 5 color themes including System (auto-follows OS dark / light mode). Status bar shows execution events in real time β€” start, done with duration, errors, and manual stops. Search & highlight logs (Ctrl+F). Right-click the task grid for processor usage and reference lookup.

AI Execution Generator

Generate and modify task flows through natural language. Multi-turn chat with LLM, expandable Processor browser with search, selective context to save tokens. One-click MCP tool description generation with smart merge. AI-powered error analysis with fix suggestions. Vision model support for image understanding via Ollama. Supports 10 LLM providers β€” only ai_provider config needed.

AI-Powered Workflow
1

Describe

Tell AI what you want in natural language β€” create new or modify existing Execution flows through multi-turn chat.

2

Generate

AI builds task flows from 80+ processor types. Browse, search, and selectively include processor context to save tokens.

3

Publish

One-click MCP tool description generation. Auto-extracts input/output schema and generates agent-friendly descriptions. Smart merge preserves existing config.

4

Auto-Fix

When execution fails, AI analyzes the error in context β€” pinpoints root cause, suggests fixes, and opens AI Assist pre-filled with the diagnosis.

All AI features activate with a single config: ai_provider. Supports 10 LLM providers including DeepSeek, Claude, Gemini, and OpenAI-compatible.

80+
Built-in processor types
10
AI/LLM providers
3
Running modes
Cron
Cron scheduling
Scenario Examples
πŸ“¦
Procurement Automation

Download sourcing request documents, merge contracts by creation date, classify by type using LLM, then send a summary email β€” all triggered by a single Claude message.

Selenium LLM classify Email
πŸ“Š
Finance Report Builder

Pull data from multiple spreadsheets via SSH/SFTP, merge and pivot, call an LLM to write the executive summary section, and export as a final Excel report.

SSH/SFTP Spreadsheet LLM summarise
πŸš€
DevOps Task Toolkit

Expose CI pipeline triggers, deployment scripts, log fetchers, and health-checks as MCP tools. Let Copilot or a chat agent coordinate complex release workflows.

HTTP API SSH exec MCP tools
πŸ§‘β€πŸ’Ό
Personal Productivity

A lightweight personal MCP tool-set on your MacBook: web scraping, local file management, calendar data extraction, and LLM-powered daily brief generation.

Local files Web scrape LLM brief
πŸ”
Web Form & Captcha Bot

Use OCR to extract text from scanned documents or screen captures, and the CAPTCHA processor to handle login gates. Combine with Selenium and GO_TO_TASK to automate form workflows end-to-end β€” retrying on failure without writing a line of code.

OCR / Captcha Selenium GO_TO_TASK
πŸ“±
On-Device AI + MCP Tools

Deploy PETP as a Docker-based MCP Tool Server on your NAS. Connect from on-device AI apps (e.g. LSnote with Gemma 4 on-device model) to call tools on demand β€” web scraping, document processing, email sending β€” all without cloud round-trips.

MCP Server On-Device AI Docker / NAS
MCP Tool Server

Every Execution becomes an MCP Tool

PETP starts an MCP server that advertises all configured executions as named tools with typed input/output schemas. Any MCP-compatible LLM client β€” Claude, Cursor, VS Code Copilot, or your own agent β€” can discover and invoke your real business workflows without any extra glue code.

MCP Server expose executions as tools
stdio transport HTTP Streaming

Request-to-Result flow

πŸ€–
MCP Client
Claude Β· Claude Code Β· Copilot
PETP self-call Β· any HTTP Streaming MCP client
β†’
πŸ”Œ
MCP Server
tools/list Β· tools/call
stdio Β· HTTP Streaming
β†’
βš™οΈ
PETP Runtime
Pipeline β†’ Execution β†’ Tasks
β†’
🧩
Processors
Selenium Β· HTTP Β· SSH Β· LLM Β· Email…
β†’
πŸ“€
JSON Result
Typed output returned to LLM
Simple Configuration

Define a complete automation in YAML β€” no code, no SDK, just declare what to do.

execution: DAILY_REPORT
astool: true
mcp_desc: '{"desc":"Generate daily report","params":["date"]}'
list:
  - type: HTTP_REQUEST
    input: '{"url":"https://api.example.com/data?date={date}"}'
  - type: AI_LLM_QANDA
    input: '{"prompt":"Summarize: {http_response}"}'
  - type: SEND_EMAIL
    input: '{"to":"team@co.com","body":"{llm_answer}"}'

Build your own MCP tool-set with PETP

Define executions in YAML, start the MCP server, and any LLM agent can immediately discover and orchestrate your real-world workflows. No SDK. No boilerplate.

Learn how it works