Install the plugin, and Claude Code builds models and memos that keep the assumptions, evidence, and math behind every conclusion.
Handing this to an agent instead of reading it? The same instructions, in plain markdown: /product/claude-code.md
DeepCell speaks two protocols to the same service. Which one you want depends on a single thing: whether your agent has a shell.
| Your agent | Use | Why |
|---|---|---|
| Claude Code, Cowork, or your own terminal | The CLI — this page | A shell means every command is available: sign-in, the Excel and PowerPoint exports, and the whole clone / pull / push / commit / merge history. |
| claude.ai, Manus, and other web agents | The MCP server | Nowhere to run a CLI. The server wraps the same commands as one tool and blocks the ones that need a browser or a working copy. |
One line to remember: shell → CLI, no shell → MCP. Setting up the MCP path instead? That lives on the Connect page.
Run these inside Claude Code. The install opens a details view where you choose the scope; the reload activates the plugin in the session you are already in.
/plugin marketplace add deepcell-ai/deepcell-plugins
/plugin install deepcell@deepcell
/reload-pluginsLater, pull new versions with:
/plugin marketplace update deepcell
/plugin update deepcell@deepcellThe plugin drives the deepcell CLI rather than bundling it. One command installs it, puts it on PATH and verifies it — on Windows, macOS or Linux, with or without pip. Ask Claude to run this for you, or run it yourself:
macOS / Linux
curl -LsSf https://beta.deepcell.net/install.sh | shWindows (PowerShell)
irm https://beta.deepcell.net/install.ps1 | iexThen check the environment in one command — it reports the version, whether the server is reachable, who you are, and which workspace is active. There is no sign-in step first: the CLI works anonymously until something needs an account.
deepcell doctorThe deepcell skill is the entry point. Claude reaches for it on its own when a request needs a model or a memo that has to hold up, and you can call it directly:
/deepcell:deepcellThen say what you want built — a goal and its constraints, not a procedure:
“Build a three-year DCF for Acme from the filings in ./filings. Keep the growth and margin assumptions where I can change them, and record why each one is what it is.”
The first thing it does is look: deepcell --help for the commands, deepcell guide for the format and the modeling. It does not guess a flag or a tag. Work big enough to own a document goes to the deepcell-builder agent, which finishes it end to end and reports back.
Come back the same way. Change an assumption and everything downstream reruns; ask what a conclusion rests on and the answer is already in the file, because it was written while the work was being formed.
| Component | Name | What it does |
|---|---|---|
| Skill | deepcell | The entry point — the same identity and quality bar the hosted DeepCell orchestrator runs on: what the work is, when it is done, and to look things up instead of guessing. |
| Agent | deepcell-builder | Owns one complete piece of work end to end and reports back. It gets a goal and its constraints, never a procedure. |
Nothing in the plugin names a flag, a tag, or a function, so nothing in it goes stale when one is renamed. The instructions own identity and judgment; deepcell guide owns the format and the modeling; deepcell --help owns the commands. The skill and the agent are generated from the same prompts the hosted agent runs on, and a drift test fails when they fall behind.
The plugin directory is also a plugin under Agent Plugins Specification v1.0.0, so clients other than Claude Code can install the same skill. Claude Code reads its own manifest and ignores the portable one; the agent above is a Claude Code concept and does not travel.