Claude Agent SDK for Python

The README explains that allowed_tools auto-approves listed tools, while permission_mode and can_use_tool decide the rest; it also describes disallowed_tools and hooks. However, the full Claude Code toolset is available by default, and examples pre-approve Bash and accept edits, with no general per-action confirmation guarantee shown. Least privilege and confirmation are therefore only partially addressed. Data flows through the Python SDK, Claude Code CLI, Claude service, and optionally external MCP subprocesses. The README outlines some architecture, but the supplied files do not describe sensitive-data retention, redaction, or access controls, so transparency and sensitive-data handling are thin. Release workflows hash-pin build dependencies, use PyPI trusted publishing, and isolate the changelog job from publishing credentials, supporting a score of 2 for dependency security. Tools can write files, run Bash, and access MCP services, while no general undo or recovery mechanism is shown; external effects and rollback are only partially addressed. README, package metadata, license, and release workflow identify the Anthropic project and update path, providing solid attribution evidence.

README, pyproject, and test configuration are broadly consistent on package identity, Python versions, async interfaces, and dependency ranges. Test files show checks for build, release tooling, and changelog structure, but static review cannot establish that tests ran, so self-consistency scores 2. Core dependencies have version bounds, the CLI is bundled in wheels, and a custom CLI path is supported; network services, platform builds, and external MCP servers remain dependencies, so availability is not fully assured. The README lists structured exception types and handling examples; failure reporting is reasonably covered, though not every runtime failure path is shown.

The SDK targets Python developers. The README covers simple queries, interactive clients, tools, hooks, working directories, prompt configuration, and internal or external MCP scenarios. It supports Python 3.10–3.13 and offers CLI path configuration, a tool allowlist, and hooks, making common scenarios and capability boundaries clear. Environment guidance is useful, but CLI compatibility requirements are specified only for some features, so these criteria earn adequate rather than full marks.

The README is organized around installation, quick start, features, types, error handling, and development and release. Prerequisites and examples for major usage patterns are clear, though FAQ and limitations coverage is sparse, especially for data governance and risk boundaries. Package names, metadata, and API concepts are consistent. The complete MIT license text merits full marks. SDK and bundled CLI versions have a clear, separate release process, and a changelog and automated release path are present. Anthropic authorship, repository identity, and publishing workflows identify maintenance responsibility, but there is no more detailed maintenance commitment, so this does not receive full marks.

The SDK provides async message streams, interactive sessions, custom tools, and hooks, with output that callers can consume. Its documented integrations offer practical value compared with invoking the CLI directly. Examples and exception types aid use. The files support the claimed added value, but do not quantify cost-benefit or sufficiently detail CLI size, network/API usage, or deployment costs, so cost-benefit scores 1.

The README links key API descriptions to source files, and configuration, dependencies, license, and release processes can be cross-checked across supplied files. Test source contains inspectable assertions for build and version constraints, but this review does not infer that they ran or passed. Documentation generally separates interface facts from examples; performance and deployment benefits remain claims without independent corroboration, so these criteria score 2 rather than full marks.