# Grasp Agents

A lightweight, strictly typed Python framework for building LLM agents and structured workflows, usable either embedded as an application's orchestration layer or standalone with cross-session memory, skills and a terminal UI.

Grasp Agents is a Python framework for LLM agents and pipelines built around static typing end to end: agents and pipeline steps are generic Processor objects typed by their input, output and shared context, and tool and agent inputs and outputs are Pydantic models. It exposes native provider adapters for OpenAI Chat Completions, Responses, Anthropic, Gemini and LiteLLM, each surfacing provider-specific settings such as reasoning effort or server-side search directly rather than flattening them to a lowest common denominator, and a FallbackLLM can cascade across providers.

Three composition patterns share one framework: a single agent driving tools, typed processor pipelines that run sequentially, in a loop or in parallel, and dynamic multi-agent graphs with per-payload routing; any of them can be wrapped as a tool for another agent. Tools, sandboxing, cross-session memory, skills and MCP support are opt-in layers, and agents and workflows checkpoint as they run so an interrupted subagent resumes without the parent re-running completed work.

A simple multi-agent terminal UI, built with Textual, is included for running agents standalone.

Website: https://github.com/grasp-technologies/grasp-agents