Open Source TypeScript Agent Harnesses

ZCode is an open-source AI coding workspace that combines desktop, browser, and terminal interfaces around a shared agent runtime. The repository includes an Electron desktop app, web client, backend services, reusable UI packages, and the ZCode Agent CLI. Its terminal interface also serves as the agent runtime used by the other clients. Shared protocols, RPC components, provider integrations, persistence services, and client SDKs keep the different interfaces synchronized. Developers can run local desktop workspaces or prepare assets for remote workspace scenarios. The project uses a pnpm monorepo with React-based UI components and centralized build tooling. It is designed as a complete coding-agent environment rather than a standalone command-line assistant.

DeepSeek Harness is an open-source agent harness built around the idea that every major capability should be replaceable as a plugin. It uses Cordis to compose model adapters, tools, session handling, agent loops, persistence, sandboxing, approvals, and interface layers into configurable profiles. Users can run a browser-based interface, select a workspace, configure DeepSeek or other compatible model endpoints, and launch agent sessions against local projects. Agents can read and edit files, execute commands, delegate work, and maintain plans. Durable session events make work replayable and support persistence, forking, and resuming. The project is currently a fast-moving developer preview, so compatibility-breaking changes are expected.

OpenClaw (formerly Clawdbot/Moltbot) is an open-source, self-hosted autonomous AI assistant designed to run on user-controlled hardware and bridge conversational natural language with real-world task execution, effectively acting as a proactive digital assistant rather than a reactive chatbot. It lets you send instructions through familiar messaging platforms like WhatsApp, Telegram, Discord, Slack, Signal, iMessage, and more, and then interprets those instructions to carry out actions such as managing calendars, sending emails or messages, browsing the web, executing system commands, and coordinating workflows across services — all while maintaining long-term memory and context across sessions. Because it runs locally or on infrastructure you choose (like a personal computer, VPS, or Raspberry Pi), OpenClaw emphasizes data ownership, privacy, and full transparency into how your instructions are handled and what actions are taken, giving users autonomy over their AI workflows.

LobeHub is an all-in-one workspace designed to help humans and AI agents collaborate, grow, and evolve together. It treats AI agents as true teammates rather than one-off tools, enabling deeper context, continuity, and productivity. Users can build personalized agent teams that understand their workflows, preferences, and goals over time. LobeHub brings multiple models, tools, and modalities into a single unified environment under the user’s control. With built-in collaboration features, agents can work in parallel, share context, and support complex projects seamlessly. The platform is built around the idea of co-evolution, where both humans and agents continuously learn and improve together.

Pi is an open-source AI agent toolkit and coding agent framework designed to help developers build, run, and extend intelligent AI-powered workflows. Developed by Earendil Works, the project includes a coding agent CLI, unified multi-provider LLM API, agent runtime, terminal UI library, and web UI components. Pi supports multiple AI providers such as OpenAI, Anthropic, and Google through a single consistent API layer, making it easier to integrate different models into applications and agents. The framework is built around extensibility, allowing developers to create custom tools, workflows, and autonomous coding experiences. Pi also encourages the sharing of real-world AI coding sessions to improve agent performance through practical usage data instead of synthetic benchmarks. With its modular architecture, active open-source community, and support for advanced agent capabilities, Pi provides a comprehensive foundation for building next-generation AI development tools and autonomous c

oh-my-agent is a flexible and extensible framework designed to simplify the creation, management, and orchestration of AI agents across various tasks and environments. It builds on the idea of modular agent systems, allowing developers to define specialized roles and capabilities that can be combined into larger workflows. The framework emphasizes usability, making it easier to configure agents, assign tasks, and manage interactions without requiring deep expertise in AI system design. It likely includes support for plugins or skills, enabling agents to extend their functionality through integrations with external tools. The system also focuses on coordination, allowing multiple agents to collaborate on complex tasks in a structured manner. Its architecture supports experimentation, making it suitable for both prototyping and iterative development. Overall, oh-my-agent provides a practical foundation for building and managing multi-agent systems.

Oh My OpenAgent is a large-scale, open-source agent orchestration framework that aims to provide a fully unified and extensible environment for AI-powered software development and automation. It builds on the idea that no single model is sufficient, instead enabling coordinated use of multiple models for reasoning, creativity, speed, and cost efficiency within a single workflow. The system is designed as a comprehensive agent harness where tasks are automatically decomposed, delegated, and executed across a network of specialized agents. It emphasizes openness and flexibility, allowing developers to integrate different providers and avoid dependency on any single ecosystem or vendor. The framework includes robust tooling for managing agent workflows, monitoring execution, and integrating external tools, making it suitable for complex, production-level use cases. It also fosters a strong community-driven development approach, with features evolving in real time.

Calabrass: Open Source Personal AI Agent Calabrass is a free, open-source personal AI assistant built to run locally on your Windows desktop. Chat with AI models, organize files in your local folders, and connect external tools via the Model Context Protocol (MCP) 🔑 Key Features 🔒 100% Private & Local: Runs entirely on your PC. No cloud account required; your data never leaves your device. 📂 Local Workspace Access: Let your agent inspect, read, and manage files directly inside your local folders. 🔌 MCP Integration: Connect external tools, custom functions, and automated workflows using MCP. 🤖 Multi-Model Support: Modern UI designed to interact with flexible AI models for coding and research. ⚡ Easy Setup: Lightweight 64-bit Windows installer with no admin rights needed. 📄 License: MIT 🌐 Official Site: https://calabrass.vercel.app/ 💻 GitHub: https://github.com/CodewithMubasher/Calabrass

Caprock is a free, open-source desktop app for AI coding agents, for macOS, Windows and Linux. It puts Claude Code, Codex and OpenCode sessions — including the ones you start in your own terminal — and the Gemini CLI sessions it starts into one window: terminals and projects in tabs and split panes, with activity, cost, plan limits and loop alerts for every session. Cost is per repository and per model, priced at API list prices from the tokens the agents already record, next to the 5-hour and weekly plan-limit windows. Everything the agent wrote is searchable across sessions and kept after Claude Code's 30-day transcript cleanup. A phone can carry on the work. It works with the subscription you already have. A daemon on loopback records sessions into a local SQLite database: no account, no hosted service, no telemetry. Free under Apache-2.0; an optional Premium licence adds paid features on top.

LazyCodex is an agent harness for using Codex on complex software projects. It is designed to add structure around AI coding sessions through memory, planning, execution, verification, skills, hooks, routing, and diagnostics. The project helps developers move beyond one-off prompts by giving the agent a more organized workflow inside a codebase. It supports project memory so context can persist across sessions and decisions do not need to be repeatedly reintroduced. LazyCodex also emphasizes verified completion, which means the workflow is built around checking whether tasks are actually finished rather than only generating code. Its main value is turning Codex into a more disciplined coding agent environment for larger and more demanding repositories.