Connect a reusable procedure to narrowly authorized external information or tools.
Conversation & research · Documented · Documentation checked 2026-10-04
Claude apps, projects, and Cowork; Claude Code and Claude Platform are separate environments.
Skills require code execution. Plugins are paid-plan capabilities; subagents and hooks differ between Chat, Cowork, and Claude Code.
The worked example is illustrative and has not been tested in a signed-in product account.
What it does and how it connects
Connect a reusable procedure to narrowly authorized external information or tools.
Skills supply task procedures and resources; connectors supply external information or actions; plugins package capabilities. MCP describes the tool interface, not a privacy guarantee. Hooks and subagents are available differently in Chat, Cowork, and Claude Code. A reusable instruction cannot enforce a service-level permission.
Walkthrough
Use the product named in the access panel. The steps below combine documented controls with the editorial example.
- Inspect the skill or plugin’s publisher, files, and requested capabilities.
- Enable code execution where a skill needs it and install the trusted package on a supported surface.
- Authorize only the intended external account and test a read-only retrieval.
- Inspect returned source identity and tool activity before adding any write capability.
Worked example
Use a lesson-writing skill with one approved source document. If a plugin provides a connector, verify that it reads the right account’s file. Drafting a lesson should not require permission to send messages or modify the source.
Follow the available lesson-writing procedure using [approved source]. Tell me which instructions and source you used. Produce a private draft and a list of factual claims to check. Read only from the specified account; do not send, share, or edit external files.
Inspect and verify the result
Confirm that the source exists in the authorized collection and the procedure does not run unexpected scripts. Review each action tool’s operation and destination. Plugin synchronization with Claude Code depends on the documented sign-in method and client version.
Limits and troubleshooting
Cloud connectors must be reachable from their supported environment; a private local service is not automatically accessible. Untrusted skills can include unsafe instructions or executable code. If a required capability is unavailable, report the dependency instead of pretending the task completed.
Data, permissions, and actions
Connectors use granted account scopes. A local MCP server may access the computer; a remote one receives requests over the network. Inspect its trust boundary and data handling. Stored project context and connected-service copies remain separate records.
Related reading
Cowork, agent tasks, and scheduling
Give an eligible agent a bounded deliverable and understand its cloud or local workspace.
From reports to an editable briefing
Build a private Artifact from a bounded source set and a verified argument.
A bounded Cowork preparation task
Use an eligible agent to prepare a workshop pack while preserving original files.
Plugins, connectors, and skills
Separate packaged capabilities, reusable instructions, and the connections that let AI reach other software.
Sources and verification
Documentation checked 2026-10-04. Examples are illustrative unless described otherwise.
Skills: instructions and resources · Anthropic
Use plugins in Claude · Anthropic