Experimental — This project is a very early preview release published for feedback and experimentation. It is not production-ready.
pera is a lightweight, embeddable execution core for agentic systems. It provides an environment where AI agents can run, with a focus on supporting code mode execution. It is designed to complement safe interpreters like Monty, but it can also work with traditional tool calling and MCP (coming soon).
The core idea is simple: extend the system by adding skills. Each skill includes:
- Instructions for the model
- API definitions
The model uses these APIs to generate code that orchestrates functionality across different domains.
pera binds function calls in generated code to their actual implementations. This ensures that:
- All function calls are defined
- Data passed between functions has the correct shape
It also enforces security constraints:
- Orchestration code runs in a safe interpreter (Monty)
- External calls/side effects happen through validated skill actions
- Skill actions run inside a lightweight sandbox with explicitly injected host capabilities
Motivation
LLMs can be faster, cheaper, and more reliable when generating Python code instead of plain text or using traditional tool calls. For more on this approach:
- Codemode (Cloudflare)
- Programmatic Tool Calling (Anthropic)
- Code Execution with MCP (Anthropic)
- Smol Agents (Hugging Face)
Inspiration