Panel: An Agentic Research Environment

Panel is a specialized workspace designed for research where an AI agent operates as a collaborator. It integrates chat interfaces, file management, PDF viewing, and notebooks into a single unified dock. Uniquely, the agent possesses the ability to construct its own custom viewers and applications on the fly to meet specific data visualization needs.

[!IMPORTANT] This project is currently in an early build phase intended for testers.

🛠️ Getting Started

Prerequisites

Before attempting installation, ensure you have the following environment:

Node.js: Version 22.18+ or 24.12+
pnpm: Installed and configured
uv: Used for Python management (requires Python 3.12+)
Claude Code: Installed and authenticated (run claude to log in)

Installation Sequence

Execute the following commands in your terminal:

# Install dependencies and sync Python environment
pnpm install
uv sync

# Launch the application
pnpm start

Once the process completes, navigate to: http://localhost:4173

Note: The initial pnpm start includes a build step, so expect a delay of about a minute during the first run.

📂 Data & Storage

To ensure your data persists even if you delete or re-clone the repository, Panel stores critical information outside the project folder:

Data Type	File Path	Description
Database	`~/Panel/panel.db`	Stores all agent interactions and conversation history.
Workspaces	`~/Panel/workspaces`	Default directory for created research workspaces.

🚀 Functional Overview

Current Capabilities

Agentic Interaction: A chat-based agent capable of reading/writing files (requests permission before tool execution).
Workspaces: Dedicated folders with isolated chat histories and saved UI layouts.
Collaborative Notebooks: Real-time editing of notebooks using a live kernel, shared between the user and agent.
Asynchronous Execution: Ability to run and monitor long-term background commands.
Dynamic Panes: The agent can program custom viewers when built-in options are insufficient.
Literature Reviews: Generate comprehensive reviews via the chat and access them through tool cards.

Current Limitations

~~Broad LLM Support~~ $\rightarrow$ Currently only fully supports Claude Code.
Hypothesis Modules: These lack dedicated views, making their output difficult to parse.

🔑 Optional OpenAI Integration

You can integrate OpenAI by copying apps/server/.env.example to apps/server/.env and defining your OPENAI_API_KEY.

What this enables:

Adds "OpenAI API" to the agent selection menu for general chat and tool usage.

What it does NOT enable:

~~Literature Reviews~~
~~Hypothesis Modules~~

These specific features require web-search capabilities, which are currently exclusive to Claude Code.

🛠️ Troubleshooting

If you encounter the error "Panel couldn't reach its server," the backend is likely offline.

Port Conflicts: If the app fails to load or a port is occupied, run: pnpm dev:doctor

🏗️ Technical Architecture

The UI Logic

The interface consists of multiple configurable windows known as Panes. These can render:

Standard files (Images, Code, Data).
Active chat sessions.
Custom extensions (e.g., a SQLite visualizer or PDB viewer) created by humans or agents.

Module Protocol

Modules function as advanced "Skills" with strict typing to facilitate complex workflows. The data flow can be represented as:

$\text{Input} \rightarrow \text{Intermediate (Observability)} \rightarrow \text{Output}$

Intermediates are critical for transparency, providing:

Chain-of-Thought logs.
Scratchpads for agentic reasoning.
Progress indicators for multi-stage, long-running jobs.

Data Abstraction Layer (DAL)

The DAL acts as a bridge between the filesystem and in-memory objects.

By mapping URIs to specific storage locations, Modules can manipulate objects without needing to manage the underlying storage logic.

License: MIT $\text{ (Open Source)}$