Skip to main content

Introduction to OpenPact

Early Release

OpenPact is in early beta and is open for testing. APIs, configuration, and features are subject to change as the project evolves. We welcome feedback at hello@openpact.ai.

OpenPact is a secure, minimal framework for running your own AI assistant. Built in Go and designed with security as the top priority, it gives you complete control over what your AI can access while keeping your data private.

OpenPact Logo

What is OpenPact?​

OpenPact is an AI orchestration framework that connects your preferred AI model to various services and capabilities through a secure, sandboxed environment. It acts as a secure bridge between:

  • AI Models (Claude, GPT, Gemini, and 75+ providers via OpenCode)
  • Communication Channels (Discord, Telegram, Slack)
  • Your Data (files, notes, calendars, GitHub issues)
  • Custom Scripts (safely sandboxed Starlark)

Think of OpenPact as your personal AI infrastructure - you control the AI, the tools it can use, and where your data goes.

Why OpenPact?​

Security First​

OpenPact implements the principle of least privilege. Your AI assistant can only use the tools you explicitly enable, and secrets never leak to the AI model.

AI Model <---> OpenPact <---> Your Services
              (sandbox)
  • Tool Allowlisting: Only explicitly configured MCP tools are available
  • Secret Redaction: API keys and tokens are automatically hidden from the AI
  • Two-User Docker Model: Container runs with separated privileges
  • No Arbitrary Code Execution: AI cannot run arbitrary commands

75+ AI Providers​

OpenPact is powered by OpenCode, giving you access to a wide range of LLM providers:

  • Anthropic (Claude), OpenAI (GPT), Google (Gemini), Ollama, and many more

Switch between models and providers without changing your configuration or losing your setup.

Sandboxed Scripting​

Extend OpenPact's capabilities with Starlark scripts - a Python-like language that runs in a secure sandbox:

  • No Filesystem Access: Scripts cannot read or write files
  • HTTP Only: Network access limited to HTTP/HTTPS
  • Execution Limits: Configurable timeouts prevent runaway scripts
  • Automatic Secret Redaction: Even script output is scanned for leaked secrets

Production Ready​

OpenPact is built for real-world deployment:

  • Docker Native: Easy deployment with official container images
  • Health Checks: Built-in endpoints for monitoring (/health, /ready, /metrics)
  • Prometheus Metrics: Export metrics for your monitoring stack
  • Structured Logging: JSON logging for log aggregation
  • Rate Limiting: Built-in request rate limiting

Key Concepts​

MCP Tools​

OpenPact uses the Model Context Protocol (MCP) to expose capabilities to AI models. Each tool has a specific purpose and clear boundaries:

CategoryTools
Workspaceworkspace_read, workspace_write, workspace_list
Memorymemory_read, memory_write
Communicationchat_send
Integrationscalendar_read, vault_*, github_*, web_fetch
Scriptingscript_run, script_exec, script_list, script_reload

Context Files​

OpenPact uses special markdown files to shape AI behavior:

  • SOUL.md: Defines the AI's identity and personality
  • USER.md: Contains user preferences and context
  • MEMORY.md: Persistent memory that the AI can read and update

Starlark Scripts​

Custom scripts let you extend OpenPact's capabilities safely:

# Example: Fetch weather data
api_key = secrets.get("WEATHER_API_KEY")
resp = http.get(format("https://api.example.com/weather?key=%s", api_key))
data = json.decode(resp["body"])

The AI sees results but never the actual API key values.

Community​