Claude for biotech: the AI that runs locally

Three surfaces carry the same name and they are not interchangeable. If you are choosing one to build a durable company record on in a small biotech, two documented differences decide it, and neither is about how clever the model is.

Anthropic ships Claude as chat, as Cowork, and as Code, and its own documentation is unusually direct that the engine underneath is the same. The docs say Cowork "uses the same agentic architecture that powers Claude Code, accessible within Claude Desktop without opening the terminal," and the accompanying tutorial says both "are Claude Code underneath, local to your machine." So the choice is not a capability ladder. It is packaging, containment, and who is expected to sit in front of it.

The three surfaces, as Anthropic describes them

On containment, the docs are equally plain. Cowork "runs in a contained space on your computer" where Claude "can read, create, and edit files within the folders you share, but can't access anything outside them," while Code "runs directly in your project with full access to your file system, terminal, and development tools."

The two facts that decide it

First: Cowork has no hooks. Anthropic's own comparison table lists hooks under Code only. That sounds like a developer detail and it is not. Hooks are how you make a system enforce its own rules rather than politely intending to follow them: refusing to state something as confirmed without evidence behind it, blocking an action that is out of bounds, forcing a check before an answer reaches you. Everything that makes an assistant safe to hand real company state to is enforcement, and enforcement lives in hooks.

Second: Cowork does not read the Claude Code directory. The docs say it loads what is enabled for your account and "doesn't read the Claude Code CLI's ~/.claude directory on your machine," and that anything living only there has to be added through the account's customise surface instead. If a system is built as a set of skills and hooks installed into that directory, moving it to Cowork is a repackaging job, not a preference.

Which is why the system this site describes is built on the enforcement surface rather than the contained one. Not because the terminal is better, but because the guardrails only exist there today, and the guardrails are the product. You do not have to live in a terminal to get them: it ships as an app, and the enforcement runs underneath it.

Why locality earns its keep in drug development

Consider what a business development lead handles in a single out-licensing process. A partner's diligence question list under a confidentiality agreement. Unpublished toxicology data. A term sheet redline with the royalty structure in it. Deal comparables. An internal deal review that says candidly which of two counterparties you would rather sign. That material is commercially sensitive before it is anything else, and most of it is covered by an agreement that constrains where it can go.

The default reflex for all of it is to paste the relevant paragraph into a chat window. What locality changes is that you do not have to. The files stay in a folder on the machine your company already issued and already secured, the record is plain text you own, and if the tooling went away tomorrow the record is still there and still readable without any particular application. That last point is not a marketing line, it is the difference between owning your records and renting access to them.

Skills, and the ecosystem around it

People searching for Claude in a biotech context often arrive looking for skills: reusable instructions that teach the assistant a specific job rather than re-explaining it every time. That is the right instinct, and it is the mechanism the system here is built out of. It is also the mechanism the vendors are converging on. Anthropic launched Claude for Life Sciences in October 2025 with collaborations announced alongside it, including one with Benchling, and Benchling has since shipped a server on the Model Context Protocol that lets outside tools query its data. The research platform and the assistant are designed to talk to each other, which means keeping both is a supported arrangement rather than a compromise, as the Benchling alternatives page sets out.

Two limits before you go further. The system described here is single-seat: one install, one owner, one mail identity, so a team is served by separate installs. And its scheduled routines need the machine to be awake, which is a real operational constraint rather than a rounding error.

For the platform comparison in more depth, read Claude Code for biotech against off-the-shelf AI tools. For what the assistant is set up to actually do day to day, Claude as an AI chief of staff, set up in a biotech is the practical version, and what an AI company brain is is the concept behind it. If you are weighing this against a general assistant instead, what ChatGPT cannot keep track of is the honest comparison, and the alternatives set covers the category incumbents.