AgentDock is a framework for building AI agents, designed to enable agents to deliver complex tasks with "configurable determinism." According to the repository materials, it consists of two main parts: AgentDock Core is an open-source, backend-first framework for building and deploying AI agents, designed to be framework-agnostic and vendor-independent, giving developers full control over agent implementation; the open-source client is a complete Next.js application that serves as a reference implementation and consumer of AgentDock Core, available at https://hub.agentdock.ai to see it in action.
The framework is built with TypeScript and emphasizes simplicity, extensibility, and configurable determinism, making it suitable for building reliable, predictable AI systems that can operate with minimal supervision. The repository notes that the framework has been retired, the prompt and documentation generator has been migrated to AgentDock/prompt-library, and the book has been migrated to AgentDock/ai-agents-book, and it suggests checking the pinned issue.
The design principles include: simplicity first, with the minimum code needed to create a usable agent; a node-based architecture, where all capabilities are implemented as nodes; tools as specialized nodes, where tools extend the node system to provide agent capabilities; configurable determinism, controlling the predictability of agent behavior; and type safety, with comprehensive use of TypeScript types.
Configurable determinism is at the core of its design philosophy, used to balance AI creativity with the predictability of system behavior. AgentNodes are inherently non-deterministic because the LLM may generate different responses each time; workflows can become more deterministic by defining the paths along which tools execute; developers can control the level of determinism by configuring which parts of the system use LLM reasoning; and even when LLM components are included, overall system behavior can remain predictable through structured tool interactions. The framework supports both deterministic workflows in traditional workflow builders (whether or not LLM reasoning is used) and AgentNodes with LLMs when adaptability is needed. Complex multi-step workflows (which may contain hundreds of deterministic steps, implemented within tools or as connected sequences of nodes) can be triggered by agent decisions, and each workflow can still execute predictably even though it is triggered by non-deterministic agent reasoning.
The core architecture is built around a modular node system, which serves as the foundation for all agent functionality. Node types include: BaseNode, the base class that establishes the core interface and capabilities of all nodes; AgentNode, a specialized core node that orchestrates LLM interactions, tool use, and agent logic; and tools and custom nodes, nodes through which developers implement agent capabilities and custom logic by extending BaseNode. These nodes interact through a managed registry and can be connected to enable complex, configurable, and potentially deterministic agent behavior and workflows.
Quick start requirements: Node.js version no lower than 20.11.0 (LTS); pnpm version no lower than 9.15.0 (required); an API key for an LLM provider (such as Anthropic, OpenAI, etc.). The steps include:Clone the repository (git clone https://github.com/AgentDock/AgentDock.git，cd AgentDock); install pnpm (corepack enable, corepack prepare pnpm@latest --activate); install dependencies (pnpm install; if a clean reinstall is needed, you can run pnpm run clean-install, which removes all node_modules and lock files and correctly reinstalls dependencies); configure the environment (create .env or .env.local based on .env.example, for example cp .env.example .env.local or cp .env.example .env, then add the API key); start the development server (pnpm dev). The open-source client can be deployed directly to Vercel. The project requires the use of pnpm and does not support npm or yarn.
AgentDock uses a BYOK (bring your own key) model: add the API key on the application settings page, or provide the key for direct API use via request headers; keys are securely stored using the built-in encryption system; the materials state that API keys are not shared or stored on the server. The API key resolution priority is: each agent's custom API key (via agent settings in the UI), global settings API key (via the UI settings page), environment variables (from .env.local or the deployment platform). Some tools also require their own API keys.
The repository contains: AgentDock Core (located in agentdock-core/), the open-source client (a complete reference implementation based on Next.js), example agents (ready-to-use agent configurations in the agents/ directory), and the AI Agents Book (guide content in the content/ai-agents-book/ directory). It can be used independently for your own application, or as a starting point for building agent-driven applications. The repository also contains multiple preconfigured agent templates, as well as example implementations.
Documentation is located at hub.agentdock.ai/docs and in the repository's /docs/ folder, including getting started guides, node creation tutorials, and integration examples; ecosystem documentation and guides are available at agentdock.ai/docs. Architectural components include: BaseNode, AgentNode, tools and custom nodes, Node Registry, Tool Registry, CoreLLM, Provider Registry, Evaluation Framework, error handling, logging, orchestration (controlling tool availability and behavior based on conversation context), sessions (managing state isolation between concurrent conversations). The storage system is currently evolving, supporting key-value storage (Memory, Redis, Vercel KV providers) and secure client-side storage, with vector storage and other backends under development.
Buildable applications include: AI-driven applications (custom chatbots for any frontend, command-line AI assistants, automated data processing pipelines, backend service integration); integration capabilities (any AI provider such as OpenAI, Anthropic, any frontend framework, any backend service, custom data sources and APIs); automation systems (data processing workflows, document analysis pipelines, automated reporting systems, task automation agents). Example agents include Dr. Gregory House (diagnostic reasoning, orchestrating multi-stage workflows with search, deep_research, and pubmed tools), Cognitive Reasoner (a multi-stage reasoning engine orchestrating seven cognitive tools: search, think, reflect, compare, critique, brainstorm, and debate), History Mentor (an immersive educational agent combining vectorized historical knowledge with search capabilities and dynamic Mermaid chart rendering), Calorie Vision (a vision-based nutritional analysis system providing macronutrient and micronutrient breakdowns from food images).
Regarding the roadmap, most improvements relate to the core framework agentdock-core, which is currently developed locally and will be released as a versioned NPM package once it reaches a stable version. The repository welcomes contributions, as detailed in CONTRIBUTING.md, and uses the MIT license. The materials mention that AgentDock Pro is coming soon, a comprehensive cloud platform including a visual workflow builder, advanced orchestration, and enterprise-grade infrastructure, where you can sign up at AgentDock.ai for early access and free platform credits at launch; because this product has not yet been released, its features and pricing details are not yet available.
