Check a package and its evidence
Source-build authoring: these agentplugins author commands require the authoring-enabled source build. The current installer release does not include them. Run relative examples from the package's parent folder and use the same exact root throughout.
Run the offline authoring loop
agentplugins author validate ./review-helper
agentplugins author inspect ./review-helper
agentplugins author test ./review-helperValidation checks the captured standard configuration and authoring readiness. Inspection shows the components and unresolved requirements. test without runtime flags checks configuration, hygiene, Skills and MCP statically. These commands do not start servers or execute a Skill's instructions.
Read the evidence layers
| Layer | What it establishes | What needs another check |
|---|---|---|
| Conformance | Captured standard data satisfies the embedded schemas/profile. | Useful behavior in a particular agent. |
| Host safety | Input meets bounded filesystem and host checks. | User authorization for installation or execution. |
| Readiness | Available evidence meets the checked policy. | Public release approval or platform acceptance. |
| Compatibility | Static support for selected targets and components. | Effective discovery and activation in a real client. |
| Toolchain | Bounded project, native-file and executable metadata. | Dependency resolution and server startup. |
| Runtime | An explicitly executed MCP check reports observed stages. | Other tools, clients, OAuth flows or untested environments. |
Read conformance and overall readiness separately. A valid manifest can still have incomplete component or host evidence. Keep “not evaluated” visible rather than converting it into a pass.
Select clients explicitly
agentplugins author compat ./review-helper --target codex,claude
agentplugins author inspect ./review-helper --target codex,claude
agentplugins author doctor ./review-helper
agentplugins author capabilitiescompat needs distinct comma-separated target IDs; it does not accept all. These static checks do not require installed clients. Doctor inspects captured native files and executable metadata without processes or network. Capabilities reports this engine's schemas, client metadata and implemented commands; it does not certify a package or establish public release qualification.
Evaluate MCP runtime explicitly
Local stdio runtime currently requires Linux, system bubblewrap and Node within the system runtime directories. Process containment (including delegated cgroup v2) and bubblewrap namespace preflights must pass before launch. See the stdio prerequisites. Streamable HTTP runtime does not use that local process sandbox and requires its own explicit network opt-in.
For a package created by the stdio tutorial, install its recognized locked dependencies separately, then select the exact server name from mcp.json:
agentplugins author test ./local-tools --runtime mcp --server local-toolsThis source command executes the selected MCP process and checks initialization and tool discovery. To call a tool, supply both --tool and --fixture. The fixture is a package-relative JSON object containing that tool's arguments. For the template's hello tool, create fixtures/hello.json containing {}:
agentplugins author test ./local-tools --runtime mcp --server local-tools \
--tool hello --fixture fixtures/hello.jsonFor Streamable HTTP, explicitly permit network access:
agentplugins author test ./docs-service --runtime mcp --server docs-service \
--allow-networkUse the real server identity from your package. The deadline defaults to ten seconds and may be set with --deadline to a positive duration of at most one minute. Inspect initialization, tools, optional tool-call and cleanup evidence individually. A runtime pass here is not execution through an installed agent, and does not prove OAuth, other tools or a public deployment.
Iterate with the source development loop
agentplugins author dev ./local-tools --server local-toolsdev runs bounded MCP cycles as the package changes; stop it with Ctrl+C. Use --once for one cycle and when requesting JSON output. Remote servers require --allow-network; tool calls still require the explicit tool/fixture pair. The command does not install into a native client.
Capture a reviewable report
agentplugins author validate ./review-helper --format json
agentplugins author inspect ./review-helper --format jsonUse --format human for human output and --no-color where needed. --include-root discloses the selected root; avoid unnecessary private paths in shared diagnostics. --release-policy adds bounded release hygiene to the package checks and is not publication approval. Skills subcommands have a smaller flag surface; check their own help.
Maintain JSON deliberately
Current source can plan deterministic normalization of one selected standard file, then write it only with explicit --write:
agentplugins author normalize ./review-helper --document plugin.json
agentplugins author normalize ./review-helper --document plugin.json --writeSelect mcp.json instead when that is the document you intend to normalize. Native import reads one explicit strict-JSON Claude MCP file; it does not search profiles. The current importer accepts only safe stdio entries containing command and optional args; HTTP/URL entries, env and extra server fields are skipped, even when the source client accepts them. Plan a new absent absolute destination before applying:
agentplugins author import native ./claude.json --from claude \
--output /absolute/path/to/new-plugin --name new-plugin \
--description 'Imported Claude MCP package'Review safe/skipped servers and unsupported fields. Add --write to publish the validated package to that absent destination. Unsafe or credential-bearing servers are skipped; inspect the result rather than assuming every original server travelled. No other native client import is exposed here.
Choose the right command boundary
Installer agentplugins validate ./review-helper reads a package for installation; agentplugins doctor diagnoses managed state. They are separate from the source-build authoring operations. Installer scope and security flags do not become authoring flags. Init creates real files; normalization and import require --write to apply their plans. Continue with the package handoff.