ワンクリックでManusで任意のスキルを実行

cao-mcp-apps

スター910

フォーク174

更新日2026年6月30日 12:20

Enable, operate, and extend CAO's MCP Apps surface — the sandboxed host-rendered fleet UI (SEP-1865) with the ui://cao/* views, the topology widget, the submit_command mutation choke point, SEP-2133 capability advertisement, and the default-off OAuth scope layer. Use whenever the user wants to turn on the MCP Apps UI, observe/steer a CAO fleet from inside an MCP App host (Claude / Claude Desktop, ChatGPT, VS Code GitHub Copilot, Microsoft 365 Copilot, Goose, Postman, MCPJam, Archestra.AI), debug why the views don't render, build the frontend bundles, or extend the surface (new view, tool, or command kind).

インストール

Codex または Claude でインストールこの Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。

Manusで実行

ソース

awslabs

awslabs/cli-agent-orchestrator

GitHub リポジトリを開く Creator のリポジトリを見る

ダウンロード

Manusで実行

関連職種SOC

SOC 職業分類に基づく

その他コンピュータ職コンピュータ・数学職·SOC 15-1299

SKILL.md

readonly

このリポジトリの他の Skills

同じリポジトリ

agui-author

awslabs/cli-agent-orchestrator

Author live dashboard UI from an agent via the `emit_ui` MCP tool. Emit one of six allow-listed components (approval_card, choice_prompt, diff_summary, progress, metric, agent_card) with JSON props and it renders in any AG-UI client watching the fleet. Use when you want the operator to see a decision, a diff, or a status readout instead of scrolling terminal text. Arbitrary HTML/markup is refused.

2026-07-17910

cao-workflow

awslabs/cli-agent-orchestrator

Author and run CAO Python workflow scripts — multi-step, parameterized, fan-out orchestrations executed by `cao workflow run`. Use when the user wants a repeatable multi-step job (e.g. data analysis over many files, a review pipeline, a parameterized batch). Authoring ends at a validated script file; running it is a separate, user-approved step.

2026-07-16910

cao-plugin

awslabs/cli-agent-orchestrator

Create a new CAO (CLI Agent Orchestrator) plugin. Use this skill whenever the user wants to add a plugin that reacts to CAO lifecycle or messaging events, scaffold a plugin package, understand plugin requirements, or integrate an external system (Discord, Slack, dashboards, logging, metrics) with CAO. Also use when the user asks what plugin events are available, how plugin discovery works, or how to install a plugin into a CAO environment.

2026-07-01910

cao-provider

awslabs/cli-agent-orchestrator

Create a new CLI agent provider for CAO (CLI Agent Orchestrator). Use this skill whenever the user wants to add support for a new CLI-based AI agent (e.g., a new coding assistant CLI), integrate a new provider, or scaffold a provider implementation. Also use when the user asks about the provider architecture, what files to modify, or how providers work in CAO.

2026-07-01910

cao-session-management

awslabs/cli-agent-orchestrator

Interact with CAO (CLI Agent Orchestrator) — launch multi-agent sessions, check status, send follow-up instructions, unblock stuck terminals, or shut down sessions. Use when working with CAO sessions in any capacity.

2026-07-01910

cao-memory

awslabs/cli-agent-orchestrator

Store, recall, and forget durable facts with CAO memory — user preferences, project conventions, decisions, and corrections that should persist across sessions and agents. Use proactively to check memory before asking the user, and to save anything worth remembering. Distinct from any provider-native memory.

2026-06-30910

name	cao-mcp-apps
description	Enable, operate, and extend CAO's MCP Apps surface — the host-rendered fleet dashboard visible inside MCP App hosts (Claude Desktop, ChatGPT, VS Code Copilot, Goose, Postman). Use when the user says "enable MCP Apps in CAO", "the ui://cao views aren't rendering", "rebuild MCP Apps bundles", "add a new ui://cao/* view", or "configure the MCP Apps OAuth scope layer". Operates on the CAO_MCP_APPS_ENABLED surface and cao_mcp_apps/ build system. Not for the localhost:9889 browser dashboard, not for plugins, providers, or session management.
compatibility	Requires CAO_MCP_APPS_ENABLED=true, cao-server running, and an MCP App-capable host (SEP-1865).

CAO MCP Apps

Operator + developer playbook for CAO's host-rendered fleet UI. Reference docs: docs/mcp-apps.md; example: examples/mcp-apps/.

Authoritative spec & sources of truth: MCP Apps Overview · Build an MCP App · capability negotiation · client matrix · stable spec 2026-01-26/apps.mdx (SEP-1865, Status: Stable) · SDK @modelcontextprotocol/ext-apps v1.7.4 (API ref · repo) · provenance PR #1865.

Turn it on

The surface is default-off. Enable and run:

export CAO_MCP_APPS_ENABLED=true
uv run cao-server        # :9889 (REST + SSE /events)
uv run cao-mcp-server    # registers tools/resources via the mcp_apps plugin

It is packaged as the built-in mcp_apps plugin (cao.plugins entry-point). The plugin's on_mcp_server hook registers the ui://cao/* resources, the five app tools, the topology widget, and advertises the io.modelcontextprotocol/ui capability — best-effort and default-off, so nothing changes when the flag is unset.

What the operator gets

ui://cao/dashboard — fleet overview + the mutation entry point.
ui://cao/agent — one terminal's status, output tail, inbox, sub-agents.
ui://cao/event-stream — live governance ticker (app-only).
cao://widget/topology + /widgets/topology/ — build-free live event view.

All mutations flow through submit_command(kind, payload) — kinds: send_message, assign, create_session (standard); interrupt, pause, resume (lifecycle); shutdown_session (destructive). For full payload schemas and scope requirements per kind, see references/submit-command-kinds.md.

Full capability scope (what the views use)

Beyond tools/call, the views exercise the spec's bidirectional channel:

Host-delegated open-link (ui/open-link) — the dashboard shows "Open full Web UI ↗" → http://127.0.0.1:9889 only when the host advertises hostCapabilities.openLinks (gate on app.canOpenLinks(); the sandbox forbids window.open).
Display modes (ui/request-display-mode) — views declare availableDisplayModes: ["inline","fullscreen"] at ui/initialize.
Streamed tool input (ui/notifications/tool-input / -partial) — render before the result lands.
Model-context notes (ui/update-model-context) — body-free gesture summaries keep the agent aware without leaking message contents.

preferredFrameSize and requiredScopes are CAO additions, not spec _meta.ui fields (the spec sizes via containerDimensions + ui/notifications/size-changed); CAO requests no elevated permissions.

See assets/mcp-apps-example.md for a worked MCP Apps integration example.

Gotchas

Host doesn't offer the views → confirm CAO_MCP_APPS_ENABLED=true and that initialize advertises io.modelcontextprotocol/ui (the host must speak SEP-1865). Non-SEP-1865 hosts still get text-only tool results.
Views are blank / fail to load → the React bundles aren't built. Run cd cao_mcp_apps && npm ci && npm run build:all. The topology widget needs no build and is the quickest smoke test (curl /widgets/topology/topology.html).
Mutations rejected with 403 → the auth layer is enabled and the token lacks cao:write/cao:admin (cao:admin for delete_session). Unset AUTH0_DOMAIN/CAO_AUTH_JWKS_URI to disable enforcement.
Events don't stream → check GET /events (SSE) directly; the bus is drop-on-slow, so a stalled consumer silently loses events — re-hydrate via cao_fetch_history.

Extending the surface

Agents emitting UI intents into this surface? Load the agui-author skill — it teaches how to call emit_ui with the six allow-listed components. Your emit_ui intents feed the L2 constructs that these views render.
Building or migrating an MCP App? Load the mcp-apps-builder skill first. It equips the official ext-apps Agent Skills (create-mcp-app, add-app-to-server, migrate-oai-app, convert-web-app) and the build guide. Use add-app-to-server when adding a new ui://cao/<name> view.
New command kind → add it to submit_command's classifier + router in mcp_server/app_tools.py (map to a real Backplane HTTP endpoint; never bypass the HTTP-only boundary) and to the scope pre-check.
New view → add a ui://cao/<name> resource in ext_apps/apps.py + an entry point under cao_mcp_apps/, build it, and tag the rendering tool with ui_meta(...). For the full step-by-step view creation procedure, see references/extending-views.md.
New host-delegated action → add a thin method on the McpApp bridge (cao_mcp_apps/src/shared/mcpApp.ts) that issues the spec ui/* request (e.g. openLink → ui/open-link, requestDisplayMode → ui/request-display-mode); gate UI on the matching hostCapabilities flag and cover it with a mockHost test.
Keep the boundary → mcp_server/* must reach state only over HTTP; the AST guard test (test/test_http_only_boundary.py) enforces it.
Keep bundles JIT-free → no eval/new Function (host CSP forbids it); the CI scan fails the build otherwise.

Recording & Verification

After building or modifying views, regenerate the demo media:

cd cao_mcp_apps && npm run build:all && npm run demo

This runs scripts/record-demo.mjs which:

Boots the E2E harness server (serves built bundles in a real MCP-host iframe)
Drives Chromium through: dashboard → agent detail → unified → event-stream
Records video (docs/media/mcp-apps-demo.webm)
Captures screenshots (docs/media/mcp-apps-{dashboard,agent,unified,event-stream}.png)
Generates an optimized GIF (docs/media/mcp-apps-demo.gif) when ffmpeg is available

The GIF is referenced in README.md and docs/mcp-apps.md — always regenerate after view changes so docs stay current.

Env overrides: CHROMIUM_BIN (path to Chrome), FFMPEG_BIN (for GIF), DEMO_PORT.

For a worked example of the full MCP Apps surface in action, see assets/mcp-apps-example.md.