AgentPilot

AgentPilot is a multi-purpose AI workflow automation system for creating, organizing, and executing complex AI-driven tasks. It supports both chatting with a single LLM and complex multi-member workflows, and provides an intuitive and feature-rich interface that lets users design AI workflows and chat in real time.

Purpose AgentPilot's core purpose is to build and manage AI workflows. Whether it is a simple single-model conversation or a complex process composed of multiple agents, code, text, prompts, and modules, it can be designed, executed, and iterated through this platform.

Main capabilities

• Create agents: You can create new agents, edit their configurations, and organize them into folders; multi-member workflows can be saved as a single agent and nested indefinitely.
• Branching workflows: Messages, tools, and code can all be edited and rerun, forming branch points, with support for cyclic switching; the branching feature applies to all plugins and multi-member chats.
• Graphical workflows: Other members or blocks can be added to a workflow, and the way they interact can be configured; vertically aligned members execute in parallel. Available member types include: User (waits for input), Agent (gets an LLM response with integrated tools and messages), Text (a text block that can nest other blocks), Code (gets code output), Prompt (gets an LLM response from a single prompt), Module (runs or gets any module's methods or variables), Workflow (any combination of the above types).
• Blocks and tools: Manage a collection of nestable blocks that can be used in any workflow or text field, enabling reuse and consistency; the default block is a text block, but it can also be any member type, or even a multi-member workflow. Blocks can be referenced by name using curly braces, such as {block-name}. Tools can be assigned to agents; by default they are a single Code member, but they can also be a complete workflow; parameters can be configured, and modifying and rerunning parameters at runtime creates branch points.
• Modules: Python files imported at runtime, suitable for scenarios that require persistence, such as toolkits, daemons, memory, and custom pages.
• Custom UI: Provides a flexible and powerful base class for building complex hierarchical configuration interfaces; the entire application is built on this framework; developers can easily modify or create configuration pages, even while the application is running.
• Scheduler (advanced): Schedule workflow runs at specific times or intervals, with support for natural language time expressions, such as every 5 minutes, every day at 3 PM, and the second Tuesday of every month.
• Structured output: Members can be configured to output structured data with the help of Instructor.
• Plugins and extensions: Custom plugins can be created and imported to extend functionality. Supported plugins include Agent (custom agent behavior), Open Interpreter, OpenAI Assistant, CrewAI Agent (currently disabled), Workflow (CrewAI Workflow currently disabled), Provider (model provider support, such as Litellm).
• Code interpreter: Integrates Open Interpreter, can be used independently as a plugin, and can also be used to execute code in 9 languages (Python, Shell, AppleScript, HTML, JavaScript, PowerShell, R, React, Ruby). Code can be executed from any Code member or message with the Code role; automatic running can be enabled for code messages in settings.
• AI generation: Blocks under the System Blocks folder are used to generate or enhance fields, and by default include Claude's prompt generator; prompts (AI-enhanced user input), agents (AI-generated agents, coming soon), system messages (AI-generated system messages, coming soon), pages (AI-generated pages, coming soon).
• Model support: Integrates LiteLLM, supports 100+ models, and supports custom API servers, Google PaLM API, Google VertexAI.

How to use Can be built from source; see "How to build from source" for specific methods. When upgrading, the old database can be migrated by replacing the executable file with the new version before starting the application. Users can create agents, organize folders, design workflows, configure tools and blocks, run code, set up scheduling (advanced), etc. through the graphical interface. Block names can be referenced in text fields using curly braces.

Use cases Suitable for scenarios requiring automation of AI-driven tasks, including single-model conversations, multi-agent collaboration, code execution, scheduled tasks, structured data output, etc. Suitable for developers and teams who want to build, organize, and execute complex AI workflows.

Limitations and notes

• The scheduler is a premium feature.
• Some features are marked as "coming soon" or "currently disabled."
• Users must always understand the code being run; executing any code is their own responsibility.
• Caution is required when using automatic code running and Open Interpreter; when opening any chat, if the last message is code, automatic running will begin.
• Windows executables must keep the console visible due to a known issue.
• Creating a venv on Linux will not install pip.
• Changing the configuration of an OpenAI Assistant will not reload the assistant; the chat must be closed and reopened.
• The project welcomes contributions; pull requests can be submitted.