Panel is an open-source research workspace whose agent builds its own UI

An agent that shares and extends the workspace

A project called Panel has surfaced on Hacker News as a Show HN post, offering a different take on how an AI assistant should sit inside a research environment. According to the project's GitHub README, Panel is a research workspace where an agent works alongside the user: chat, files, PDFs and notebooks live in a single docked interface, and when a built-in viewer cannot display something, the agent can write a custom pane or app to show it. The developers describe it as an early build for testers and ask users to report issues.

Panes, workspaces and shared notebooks

The interface is built from configurable windows called Panes, which can show images, data files, code or chat sessions — a layout aimed at researchers who constantly switch between file types. A default set covers common cases, but custom Panes can be added by either the human or the agent; the README's examples include a viewer for PDB molecular structure files and a SQLite browser.

Each Workspace is a folder the agent operates in, with its own conversations and a saved layout. Conversations and everything the agents did are stored in ~/Panel/panel.db, while new Workspaces default to ~/Panel/workspaces unless another folder is chosen. Both live outside the repository, so deleting or re-cloning the code leaves the data intact.

The agent can read and write files, but asks for confirmation before running any tool. Jupyter notebooks run against a real kernel, and the user and the agent can edit the same notebook. Long-running commands can be pushed to the background, where they can be watched and stopped. There is also a literature review feature: request one in chat and open the result from a tool card.

A typed module protocol and a data layer

Under the UI, the README describes two structural ideas. The first is the Module Protocol. Modules are described as akin to Skills, but with typed definitions for Inputs, Outputs and Intermediates. Intermediates are objects that surface observability — the chain-of-thought or scratchpad of an agentic Module, or intermediate outputs in a multi-stage Module — which matters for processes that need transparency or long-running jobs that should show progress. The typed definitions enable validation at runtime and are intended to make it easier for humans and agents to build custom Modules and the Panes that visualise them.

The second is a Data Abstraction Layer that bridges in-memory and filesystem objects. It maps a URI to either an in-memory store or a local file, so a Module only has to manipulate the object rather than worry about where it lives.

Setup and current limits

Running Panel requires Node 22.18 or newer (or 24.12 and above), pnpm, and uv, which fetches its own Python 3.12 or newer. Claude Code must be installed and signed in, because both the agent and the literature review run through it. Setup is pnpm install, uv sync and pnpm start, after which the app is served at localhost:4173. A pnpm dev:doctor command helps diagnose port conflicts and similar startup problems. The project is MIT-licensed.

Limitations are stated openly: only Claude Code is fully supported; Modules can be started only by asking in chat, with no launch button; hypothesis Modules have no view of their own, so their results can be hard to read; and Modules do not yet work with the OpenAI API. An optional OpenAI key adds an "OpenAI API" agent for chat and tools, but literature reviews and hypothesis Modules need a web-search-capable agent, and today only Claude Code qualifies.

Why it matters

Most AI-enabled workspaces confine the agent to a chat sidebar and hand it a fixed toolkit. Panel inverts that: the interface itself is programmable by the agent, so the workspace can grow a purpose-built viewer at exactly the moment a task needs one. The typed Module protocol, with explicit Intermediates for scratchpads and progress reporting, is a credible attempt at observability for agentic pipelines — a pain point in most agent tooling. It remains rough, early and tied to a single agent platform, but as a sketch of a workspace that the human and the agent build together, it points somewhere interesting.