SUMM

Presenter introduces themselves as an engineer on the cloud code team and outlines the agenda: SDK quick start, live GitHub action demo, advanced SDK features, and setting up the GitHub action.
Asks audience about their familiarity with cloud code and claude.
Encourages installing cloud code locally to follow along.

Cloud code SDK enables programmatic access to the cloud code agent in headless mode, acting as a foundational building block for new application types.
The SDK can be used like a Unix tool, integrated with bash or terminal workflows, and used in pipelines.
Supports use cases like code review, automated linter creation, and building custom chatbots or remote code writing systems.
Upcoming support for Python and TypeScript bindings.

Calling claude via claude-b <prompt> executes tasks (e.g., generating a Fibonacci sequence generator).
Using the d-allow tools write flag grants permission for writing files.
Examples shown: piping log files to Claude for summarization and interpreting outputs like ifconfig.
Output can be requested in JSON format for programmatic parsing and building downstream integrations.

Live demonstration using a small open source quiz app repository.
Claude can be prompted via GitHub issues to implement new features (e.g., '50/50' powerup and 'skip question' options), even with sparse specifications.
Additional issue (per-question timer) is also submitted for Claude to build.
Claude's workflow: responds to the issue, creates a job run with logs, generates a to-do list, and works through the implementation.

Claude can be asked to review or modify pull requests (e.g., changing background color from blue to green).
Claude adds new commits to already open pull requests, fully automating simple code changes on request.

By default, cloud-ba has no edit or destructive permissions; permissions must be explicitly granted via the d-allowed tools flag.
Permissions example: granting bash access to npm run build, npm test, and file writing.
Structured output available in "stream JSON" (chunked updates) or "JSON" (final output only) modes for easier downstream processing.
System prompts can be customized (e.g., making Claude respond "like a pirate").
Session IDs enable interactive, stateful experiences where the context is preserved across exchanges.
Recent addition: d-permission prompt tool enables real-time, user-driven permission management via an MCP server, removing need to pre-declare tools.

Claude works through the power-up feature request, creates new code and a pull request.
Demonstrated testing of the new feature directly in the app, with live changes reflected (e.g., new 50/50 and skip question buttons).
Claude successfully executes additional code changes as requested (color changes), updating all relevant files.

Claude GitHub action can read code, create PRs from issues, add commits to existing PRs, answer technical questions, and review code—all via GitHub comments.
No infrastructure burden: action runs on GitHub-hosted runners, leveraging existing CI infrastructure.

GitHub actions are built on top of the SDK.
Two main open source actions: base action (talks directly to Claude), and PR action (handles advanced PR/comment features like to-dos and rendering).
The codebase is open source for reviewing and building further customizations.

To install, open Cloud Code in the target repo and run /install github action from the terminal.
Guided flow walks the user through setup, resulting in a PR for the YAML workflow file.
Bedrock and Vertex users follow slightly different manual installation steps, detailed in the docs.

Links to the open source repos for the base and PR actions are provided.
Users are encouraged to give feedback via the public GitHub repo by filing issues.
Presentation concludes with thanks and an invitation to engage further.

Building headless automation with Claude Code