🇺🇸 USA · Claude Cowork

Status: 🟩 COMPLETE 🟦 LIVING Last updated: 2026-06-26 Plain-English tagline: A collaborative mode where multiple Claude Code instances work alongside you on the same project — like having a small team of AI collaborators each focused on a different angle, with you orchestrating.

Front-matter facts

Field	Value
Vendor	Anthropic (San Francisco, USA)
Country / origin	🇺🇸 USA
Recommended for Australian users?	✅ Yes — same Claude infrastructure
Privacy summary	Same posture as Claude Code overall; opt-out applies if relevant
Free tier	Limited — Cowork features available at Pro and above
Paid tiers	Bundled with Claude Pro / Max / Team / Enterprise — usage limits vary by tier
First released	2025 — phased rollout through Claude Code
Last reviewed	2026-06-26
Official site	https://docs.claude.com/en/docs/claude-code/cowork

What it is

Claude Cowork is a multi-instance collaborative mode in Claude Code. Instead of having one Claude conversation working on your project, Cowork lets multiple Claude instances work in parallel, each on different sub-tasks or different angles, with results coordinated back to you.

The mental model:

You = the orchestrator / decision-maker
Cowork agents = focused Claude instances, each working on a piece in parallel
Outputs = surfaced back to you for review, merging, or further direction

Practical use:

“Three Cowork agents — one writes the API code, one writes the tests, one writes the docs. Coordinate.”
“While I’m reviewing PR #123, Cowork agent 1 explores how feature X works in the codebase, Cowork agent 2 drafts the migration plan, Cowork agent 3 watches for any conflicting open PRs.”
“Spin up 5 Cowork agents to evaluate 5 different design approaches in parallel, then summarise the tradeoffs.”

Cowork is distinct from Subagents (which Claude Code spawns automatically for specific scoped tasks like Explore / Plan / code-reviewer). Subagents are tactical and Claude-initiated; Cowork is strategic and user-orchestrated.

Analogy: if your normal Claude Code session is “1:1 coding with a colleague,” Cowork is “running a small focused team meeting where you delegate parallel research and synthesis.”

What you’d use it for

Parallel exploration — many angles on the same problem at once
Long-running background research — while you continue coding, Cowork agents investigate adjacent concerns
Multi-dimensional design evaluation — agent per design option
Diff-and-conquer refactoring — agent per file or per module
Cross-codebase consistency — agents investigate how a pattern is used in different parts of the repo simultaneously
Documentation while coding — one agent codes, another documents the API as it lands
Test-while-build — one agent implements, another writes tests against the spec

How to use it

Activate Cowork in Claude Code

Open Claude Code in your project
Type a slash command like /cowork to enter Cowork mode (exact command name may vary; check current docs)
Define the parallel tasks: “Agent 1: explore the auth module. Agent 2: investigate the email module. Agent 3: summarise the test coverage.”
Claude spins up the requested instances; each runs in parallel
Output streams back as each agent makes progress

Coordinate output

Each Cowork agent reports back at completion
You can ask the main Claude conversation to synthesise the parallel outputs
Or use Cowork agents for fire-and-forget tasks while continuing other work

Limits and quotas

Number of concurrent Cowork agents depends on your Claude Code subscription tier
Heavy usage counts against Claude Pro / Max / Team API quotas

How it compares to alternatives

Capability	Claude Cowork	OpenAI Codex (cloud)	Cursor Composer	Devin
Multi-instance parallel	Yes (designed for it)	One task at a time	Single agent	Single agent
User-orchestrated	Yes	Auto-planned by Codex	Composer plans	Devin plans
Background tasks	Yes	Yes	Limited	Yes
Integration with main Claude session	Native	Separate from ChatGPT	Same Cursor app	Separate product

Cowork’s distinctive design choice is user-driven parallelism — you decide what runs concurrently, rather than the AI deciding sub-task decomposition.

Privacy / data handling

Same as Claude Code overall — opt-out of training applies if you’ve set it; Team / Enterprise tenant-isolated
Each Cowork agent inherits the same data-handling posture as your main Claude session
More parallel agents = more concurrent API usage = higher cost; monitor billing

Recent changes

2026: Cowork expanded and refined; coordination UX improved
Late 2025: Initial Cowork rollout in Claude Code

Gotchas

Cowork agents don’t share context with each other directly — they share context with YOU (the orchestrator); coordination across agents must be mediated through you
Easy to spin up too many agents — context costs compound. Use Cowork when parallelism truly helps, not for the sake of using it
Long-running Cowork sessions consume API quota fast — watch usage at heavy parallelism
The Cowork name is sometimes confused with team / collaboration features in general — it’s specifically the multi-instance mode
MCP servers can be configured per-agent or shared — verify which behaviour you want
Subagents (Explore, Plan, code-reviewer) ≠ Cowork agents — subagents are Claude-spawned; Cowork agents are user-spawned
For simple delegation use cases, subagents are usually enough — Cowork is for genuinely parallel orchestration

Tech & AI, Explained

Explorer

claude-cowork