copy_object¶

Copy a cloud object, or every object under a prefix, to another location. Write tool: removed by --mcp-read-only, and hidden from a chat profile without Allow writes.

Parameters¶

Name	Type	Required	Description
`from`	string	yes	Source cloud URL. Ending in `/` copies the whole folder recursively.
`to`	string	yes	Destination cloud URL. For a folder source this is the target folder.

URL schemes: s3://bucket/key, az://container/key, gs://bucket/key.

Behaviour¶

Source and destination may be different buckets, different accounts, or different providers. Which path runs is decided from the URLs:

Same provider and bucket: the backend copies server-side. No bytes travel through the server, so object size costs nothing.
Anything else: the object is streamed in 8 MiB blocks into a multipart upload on the destination. Memory stays flat regardless of object size.

The source is never modified. A folder's shape is recreated under the target prefix (data/nested/b.csv → backup/nested/b.csv).

Limits and refusals¶

More than 10,000 objects in one call is refused rather than partly done.
Copying a folder into itself or a descendant is refused.
Copying an object onto itself is refused.

Response¶

{
  "from": "s3://a/data/",
  "to": "gs://b/backup/",
  "objects": 42,
  "bytes": 10485760,
  "server_side": false
}

server_side tells you which path ran.

Credentials¶

The MCP server uses the ambient chain (AWS_ env, a cached SSO session, Azure CLI login, Google application-default credentials); the in-app assistant uses your saved cloud connections and additionally requires Allow writing to cloud storage plus that connection's Allow writes*.