Security
Sume CLI is designed for agent use, but CLI output can still contain sensitive or user-owned data. Treat all auth, upload, job, and media output carefully.
Secrets
Never print or commit:
SUME_API_KEY;- local
~/.sume/config.json; - login URLs or user codes;
- signed upload URLs;
- storage object keys;
- upload headers;
- raw provider payloads.
Use environment variables or secret managers for automation:
Local config
The default config file is:
Use an isolated config directory for disposable environments:
Before snapshotting a reusable sandbox:
Signed upload URLs
sume uploads presign returns a temporary storage upload_url. Treat it as a
secret-like value.
When uploading bytes:
- send only the returned
required_headers; - do not send the Sume API
Authorizationheader to storage; - do not log the full
upload_url; - do not log
object_keyunless engineering specifically asks for it.
Media URLs
Sume job and asset results may include public media URLs scoped to Sume output. They are still user data.
When reporting results:
- summarize media counts and file types;
- use local downloaded filenames rather than full remote URLs;
- redact query strings and private identifiers;
- avoid dumping large raw result payloads.
Paid/provider operations
These commands can spend credits or create provider work:
sume images generatesume videos generatesume ads videos createsume face-swap createsume reference-analysis createsume assets finalizewhen it starts Asset Library processing
Agents should:
- confirm user intent;
- check
sume credits --json; - run one bounded job first when testing;
- stop on API/provider errors instead of retrying blindly;
- report job ids and sanitized error codes, not raw provider payloads.
MCP gates
Default MCP is read-only:
Mutating non-paid tools require:
Paid/provider tools require:
Only enable write or paid gates in MCP hosts that have an explicit human approval model.
Agent reporting rules
Agent final reports should include:
- what command was run;
- whether it was read-only, write, or paid/provider;
- sanitized status/result summary;
- job ids only when useful for follow-up;
- local output paths if files were downloaded.
Agent final reports should not include:
- API keys;
- live login URLs/codes;
- signed upload URLs;
- storage object keys;
- raw emails;
- workspace/user ids;
- full private media URLs;
- raw provider payloads.