For AI agents: the complete documentation index is available at https://scriptedalchemy.github.io/agent-bundle/llms.txt, the full documentation bundle is available at https://scriptedalchemy.github.io/agent-bundle/llms-full.txt, and this page is available as Markdown at https://scriptedalchemy.github.io/agent-bundle/examples/index.md.
Skip to main content
  • English
  • Examples

    The repository ships seven examples, and they are products rather than fixtures. Each one uses only public agent-bundle exports and workspace:* dependencies, each one builds and validates through the same public CLI you would use, and none of them needs an API key or a signed-in host to reach a useful state. Four have a walkthrough here:

    ExampleWhat it provesRun
    Skills StarterSkills discovered by convention, with deterministic eval evidence.pnpm example:skills
    Hooks and ScriptsThe generated hook wrapper and script process envelope, including a blocking exit code.pnpm example:hooks
    MCP AppA generated stdio MCP server plus an interactive MCP App resource.pnpm example:mcp-app
    Audiobook CuratorA complete route-tree application: MCP, projected CLI, providers, durable state.pnpm example:audiobook

    Three more are advanced composition references, documented by their own READMEs rather than walked through here:

    ExampleWhat it proves
    Worktree ProximityOne root task and two child agents in linked worktrees, coordinated through durable notices.
    RSC Agent RuntimeA public example: one RSC runtime shared by hooks, MCP tools, and an MCP App timeline. The development session uses development entries and HMR topology while compiling production-compatible Flight and App output. Browser surfaces target Chromium 144+ (Claude Desktop, Cursor, Codex). Not a framework public API.
    Host TestA probe plugin that records every hook payload and MCP call a host delivers, with the request.lineage each one resolved to. Its probe:install / probe:capture / probe:uninstall scripts drive a real host in an isolated home; probe:capture claude --scenario <file.json> runs an ordered list of prompts as one multi-turn Claude Code session (claude -p --output-format stream-json, later turns --resumed), saving the model's own tool-use stream beside the hook log; the Codex and Cursor drivers take a scenario's first turn only and refuse longer ones. The plugin's dump MCP tool returns the newest 50 matching records unless limit is given (matched counts the whole log). The Workbench walkthrough needs no signed-in host; the probe lifecycle does.

    Running them

    Run the pnpm example:* scripts from the repository root. Each one builds the workspace first, then starts that example's foreground agent-bundle dev server and prints the loopback Workbench URL. Pass --open (for example pnpm example:<name> -- --open) to open a browser:

    pnpm install
    pnpm example:skills

    For the noninteractive path, pnpm examples:check runs every example package's own check script, validate and build, plus typecheck and tests where the example has them, with no dev server and no browser:

    pnpm examples:check

    Each example package also exposes the same public commands directly, once the repository-level pnpm build has built the local agent-bundle workspace dependency. Every example has validate, build, and check; the four walkthrough examples, Worktree Proximity, and Host Test also have dev, while the RSC Agent Runtime demo is driven through its tests and eval:hosts instead:

    cd examples/skills-starter
    pnpm validate   # agent-bundle validate
    pnpm build      # agent-bundle build
    pnpm check      # validate and build, without opening the Workbench
    pnpm dev        # agent-bundle dev

    Reading an example

    Every example page below records the same four things: what the example proves, the public packages it depends on, its root run command, and a link to its source. Read the source next to the page, the configuration files are short on purpose, because the src/ conventions carry most of the structure.

    Two habits make the Workbench trustworthy while you explore:

    • Wait for a completed state. A rebuild is finished when it reports Idle or Failed. A Building state is still in progress, and judging one is how a healthy project comes to look broken.
    • Treat a failed rebuild as last-good behavior. A failed build publishes no epoch, so the Workbench keeps serving the previously published artifact alongside the new diagnostic. Restore the checked-in source and rebuild; the new active epoch is the repair evidence.

    Hooks and Scripts ships a scripted, reversible diagnostic walkthrough for exactly that reason, break a handler, watch the failed rebuild keep the last-good artifact, restore, and read the new epoch as the repair. Skills Starter practices the same rebuild-and-restore loop against stale eval evidence; the other walkthroughs only ask you to wait for a completed state before judging a rebuild.

    Where the concepts live

    Example surfaceGuide page
    Skill documents and their references/ and assets/Skills
    Lifecycle hooksHooks
    Scripts and asset copyingScripts and assets
    MCP servers and MCP AppsMCP servers and MCP Apps
    The bin/lib package buildCLI and library package entries
    Eval suites, cases, and harnessesEvaluations