Directory / Claude Agent SDK

Claude Agent SDK

Agent framework · Anthropic · 6/7 verified · Last checked 2026-10-16

Shares Claude Code's permission system. canUseTool handles approvals for calls not already allowed, and PreToolUse hooks run on every call.

Overall: B (2.33 avg, 6/7 controls verified)

Lock this down first

Put must-hold checks in a PreToolUse hook, not in canUseTool, which skips pre-approved calls.

Criterion	Rating	Notes
Permission granularity	2/3	Allow, deny and ask rules plus canUseTool. We found no SDK doc proving admins can lock them, so 2.
Human approval gates	3/3	canUseTool callback asks your code to approve calls that aren't pre-approved.
Sandboxing	2/3	Same Bash sandbox as Claude Code, set programmatically.
Secrets handling	Unverified	Not confirmed in the docs we checked.
Audit logs	2/3	OpenTelemetry metrics and events, opt-in.
Kill switch / interrupt	3/3	AbortController cancels a query; interrupt() in streaming mode.
Network egress control	2/3	Sandbox network allowlist, including strict mode.

Unverified: Secrets handling. That means we could not confirm it in public docs, not that it is missing.

Vendor? Request a correction. We review corrections within 7 days.