Open Source Python Agent Harnesses for Mac

Claw Code is an open-source AI agent harness project focused on building better tools for orchestrating and managing autonomous coding agents. It originated as a clean-room reimplementation inspired by the architecture of Claude Code, aiming to replicate core concepts without using proprietary code. The project provides a Python-based foundation for experimenting with agent workflows, tool integration, and task execution pipelines. It emphasizes harness engineering—how agents are structured, how they interact with tools, and how they maintain context during execution. The system is being actively expanded, with a Rust-based runtime in development to improve performance and memory safety. Overall, Claw Code serves as a research-driven platform for advancing agent-based software development systems.

Raven is a pre-alpha multi-agent orchestration system that coordinates specialized agents through one persistent workspace. It ships with built-in agents for research, coding, visual design, and supervision of long-running jobs. Raven can also orchestrate third-party agents such as Claude Code, Codex, OpenCode, GitHub Copilot, Pi, and others. Its runtime is powered by EverOS, allowing user context, agent experience, and reusable knowledge to persist across sessions. The system can generate task graphs, delegate work, combine results, and evolve tools, skills, and workflows over time. Raven supports a terminal interface, web UI, plugins, provider routing, and self-hosted deployment. It is designed for complex, long-horizon work that benefits from collaboration among agents with different strengths.

Desloppify is a utility-focused project aimed at improving the quality, structure, and clarity of generated or written text by removing redundancy, noise, and unnecessary verbosity. It is designed to “clean up” outputs, particularly those produced by AI systems, making them more concise, readable, and professional. The system likely applies heuristics or transformation rules to identify repetitive patterns, filler content, and stylistic inconsistencies. This makes it especially useful in workflows where AI-generated text needs to be refined before publication or use in production. It may also support integration into pipelines, allowing automatic post-processing of outputs. The project reflects a growing need to manage and optimize AI-generated content rather than simply produce it. Overall, desloppify acts as a refinement layer that enhances clarity and usability of textual outputs.

The Griptape framework provides developers with the ability to create AI systems that operate across two dimensions: predictability and creativity. For predictability, Griptape enforces structures like sequential pipelines, DAG-based workflows, and long-term memory. To facilitate creativity, Griptape safely prompts LLMs with tools (keeping output data off prompt by using short-term memory), which connects them to external APIs and data stores. The framework allows developers to transition between those two dimensions effortlessly based on their use case. Griptape not only helps developers harness the potential of LLMs but also enforces trust boundaries, schema validation, and tool activity-level permissions. By doing so, Griptape maximizes LLMs’ reasoning while adhering to strict policies regarding their capabilities.

Deep Agents is an open-source, batteries-included agent harness designed for long-running, multi-step AI work. It provides an opinionated agent setup while allowing developers to override or replace individual pieces without forking the project. The framework is model-agnostic and works with tool-calling models from hosted providers, open-weight deployments, or local runtimes. Built on LangGraph, it includes persistence, checkpointing, streaming, and production-oriented orchestration. Agents can delegate to sub-agents, work with files, run shell commands, manage long contexts, and retain memory across sessions. It also supports human approval of tool calls, reusable skills, custom tools, and MCP servers.

Given the remarkable abilities of recent Large Language Models (LLMs), there is an unprecedented opportunity to build intelligent applications powered by this transformative technology. The top question for any enterprise is: how best to harness the power of LLMs for complex applications? For technical and practical reasons, building LLM-powered applications is not as simple as throwing a task at an LLM system and expecting it to do it. Effectively leveraging LLMs at scale requires a principled programming framework. In particular, there is often a need to maintain multiple LLM conversations, each instructed in different ways, and "responsible" for different aspects of a task.

FrontierAgent is an open-source agent runtime, terminal application, and evaluation suite for long-horizon research and file-based work. Its ReAct workflow provides a stateful agent that can research, read files, execute commands, and create deliverables inside a task-scoped sandbox. Agent Team mode adds a coordinator that divides work among parallel sub-agents and synthesizes their results. A live task board tracks pending, active, completed, blocked, and canceled work. Users can provide new instructions while an agent is running without discarding the active session. The same modular workflow engine also powers benchmark evaluations and can connect to OpenAI-compatible model endpoints.

GELab-Zero is an open-source “GUI Agent” framework aiming to automate interactions with graphical user interfaces (GUIs), combining both the agent model and all supporting infrastructure — including inference, input orchestration, and GUI automation logic — in a plug-and-play package that runs locally, without cloud dependencies. The idea is to let developers or users harness an AI agent that can simulate clicking, typing, reading UI elements, and interacting with apps in a human-like way via the GUI, which can enable tasks like automated testing, scriptable workflows, or even autonomous usage of GUI-based applications. Because GELab-Zero is fully open-source and doesn’t require external services, it offers privacy and control: everything runs locally under your control. The project provides a lightweight base model (4B parameters in its public release) that can run on modest hardware (depending on quantization), making it more accessible than many large-scale AI solutions.

Harness-1 is a 20B search agent trained with reinforcement learning inside a stateful retrieval harness. It is designed for long-horizon search tasks where the model must search, inspect documents, curate evidence, verify claims, and decide when enough evidence has been gathered. The harness externalizes search state, including candidate documents, evidence links, verification records, and budget-aware context. This lets the policy focus on higher-level decisions instead of trying to keep every detail inside the model context. The repository includes inference utilities, training scripts, evaluation runners, dataset tools, and documentation for running the released checkpoint. Its main value is showing how a smaller open model can approach advanced search-agent behavior through structured retrieval state and reinforcement learning.

OSS-Fuzz-Gen is a companion project that helps automatically create or improve fuzz targets for open-source codebases, aiming to increase coverage in OSS-Fuzz with minimal maintainer effort. It analyses a library’s APIs, examples, and tests to propose harnesses that exercise parsers, decoders, or protocol handlers—precisely the code where fuzzing pays off. The system integrates with modern LLM-assisted workflows to draft harness code and then iterates based on build errors or low coverage signals. Importantly, it aligns with OSS-Fuzz conventions, generating corpus seeds, build rules, and sanitizer settings so projects can plug in quickly. Reports highlight what functions were targeted, how coverage evolved, and where manual hints could unlock more paths. The goal is pragmatic: shrink the gap between “we should fuzz this” and “we have robust fuzzing running in CI,” especially for understaffed maintainers.