Agent notebook

An agent notebook is more than a notebook with a chat panel. It is a document that understands its own structure, knows what data and code it contains, and can ask a coding agent to change, extend, or repair itself. The agent sees the notebook's context, makes edits, and returns a diff that you accept or reject.

From chat to structured delegation

Most AI coding assistants today run in a chat window next to your editor. That is useful for quick questions, but it breaks down when the task spans multiple files, data sources, or reasoning steps. The chat becomes a wall of text, and you are left copying suggestions by hand.

An agent notebook fixes this by making the document itself the context. The notebook already contains your imports, queries, charts, and notes. When you ask an agent to add a new analysis, it can read the existing cells, understand the schema, and insert a new cell in the right place. The work is not a suggestion in a chat pane; it is a concrete change inside the document.

SpurLab's agent notebook

SpurLab's agent notebook is built on the Agent Client Protocol (ACP). ACP is a JSON-RPC over stdio protocol that lets any compatible coding agent connect to a host and receive tasks. In SpurLab, the notebook is the host. You type a natural-language request, tag a worker with @worker:claude or @worker:codex, and the agent receives the notebook context plus your instructions.

The agent works in a sandboxed environment. It can read the notebook, query connected data, and propose edits. Because SpurLab runs locally, your source code and data stay on your machine. The agent uses your own API keys, so you remain in control of costs and credentials.

Every agent action is recorded as a bead, SpurLab's unit of change. You can review the bead, see what files or cells changed, and roll back if the result is wrong. The notebook becomes a reproducible log of human intent and agent execution, not a black box.

Why this matters for product teams

Product teams live in documents, spreadsheets, dashboards, and code. An agent notebook can unify those surfaces. You can write a product spec in markdown, connect a usage table, ask an agent to generate a forecast model, and then ask another agent to draft the implementation ticket. The whole flow stays in one place, versioned and reviewable.

FAQ

Which agents can I use inside a SpurLab notebook?

Any ACP-speaking agent, including Claude Code, Codex, Gemini, Kimi, and OpenCode. You can also add custom agents that expose the same protocol.

Does the agent have access to all my files?

Only the workspace and permissions you give it. SpurLab keeps the notebook and connected apps on your local disk. The agent sees the context you delegate, not your entire drive.

Can I undo an agent change?

Yes. Every agent action is stored as a bead. You can review, accept, reject, or roll back to any previous bead in the notebook history.

Do I need to write code to set up an agent notebook?

No. SpurLab ships with the agent notebook ready. Install an app, connect your keys, and start typing requests.