Claude Cowork vs Claude Code

An exhaustive, source-verified comparison of Anthropic's two agentic products, written as instructor material for an executive training curriculum.

Compiled 2 August 2026 · Every factual claim traces to an official Anthropic source (claude.com/docs, code.claude.com/docs, support.claude.com, claude.com product & blog pages). Full citation list in §16 Sources.

Audience note

This pack is written for you, the instructor — not for the executives. It is deliberately more detailed, more hedged, and more citation-heavy than anything you would put in front of a class. Slide-ready simplifications and the actual curriculum outline live in §11 Teaching kit.

The one-sentence answer

Cowork

An agentic workspace for knowledge work, living inside the Claude app (desktop, web, mobile). You describe an outcome; Claude works across your files, connectors and browser and hands back finished deliverables.

Claude Code

An agentic coding tool, native to the terminal and the developer's toolchain. It reads a codebase, edits files, runs commands, and integrates with git, CI/CD and IDEs.

The critical framing for an executive audience: these are not competitors, and they are not two versions of the same thing. Anthropic's own documentation states that Cowork "uses the same agentic architecture that powers Claude Code, accessible within Claude Desktop without opening the terminal." One engine, two products, shaped around two different kinds of work and two different kinds of user.

How this pack is organised

Executive summary

The ten things that actually matter, plus the differences that are commonly misunderstood.

What each product is

Official definitions, positioning, target users, and the shared lineage.

Architecture & execution

The agentic loop, where code actually runs, sandboxing, sessions and context.

Surfaces & availability

Every place each product runs, and what you lose on each surface.

Capability matrix

The long side-by-side table. The reference section of this pack.

Config & extensibility

CLAUDE.md vs Projects, skills, plugins, MCP/connectors, hooks.

Autonomy & permissions

Permission modes, background work, Dispatch, scheduling, multi-agent.

Security & governance

Threat model, admin controls, telemetry, data retention, compliance gaps.

Plans & cost model

Which plan unlocks what, and why Cowork burns limits faster.

Which one, when

Decision tree, role-based recommendations, anti-patterns.

Teaching kit

Module plan, demos, exercises, FAQ, and the misconceptions to pre-empt.

Adoption frameworks

Anthropic's five-level maturity model, six-month roadmap, governance postures, customer outcomes.

Anthropic's training material

What to assign as pre-work, what to avoid, and the gaps you have to fill yourself.

Concept toolkit

The categorical tools worth keeping, the axis that unifies four separate frameworks, and three that don't exist yet.

Figure index

Every illustration in the pack, with what it shows and where it's used.

Sources

Every URL used, what was taken from it, and what could not be verified.

Reading conventions used throughout

Mark	Meaning
✓	Documented as available.
◐	Partial, conditional, beta, or research preview. The condition is always stated.
✗	Documented as not available, or absent from the documentation where its counterpart is explicitly present.
`—`	Not applicable to that product's shape.
Cowork Claude Code	Product tags used inline where a claim applies to only one product.

Volatility warning — read before teaching

Both products ship weekly. Claude Code publishes a weekly "what's new" digest; the Claude Desktop changelog (which covers Cowork) shows multiple releases per week. Several features described here are explicitly labelled beta (Cowork on web and mobile; Cowork scheduled tasks) or research preview (computer use; Claude Code's agent view), and Cowork's admin/compliance surface is visibly still filling in.

Practical rule: re-verify §04, §05, §08, §09 and the course catalogue in §13 before each cohort. Everything in §02 and §03 — the conceptual model — is stable and safe to teach for longer.

What this pack deliberately does not do

No third-party benchmarks or reviews. You asked for official sources only; every claim here comes from Anthropic-published material.
No pricing guarantees. Prices in §09 are as published on claude.com/product/cowork on the compile date. Treat them as illustrative in class.
No feature-completeness claim for Claude Code. Claude Code's documentation runs to over 200 pages; this pack covers the dimensions that matter for a comparison, not every flag and setting.