AI Marketplace 发布：跨 Codex、Cursor、Copilot、Claude Code 管理版本化 AI 包

AI Marketplace discovers, installs, updates, migrates, hotloads, offloads, and removes versioned AI packages from GitHub, Azure DevOps, and GitLab repositories.

Supported AI vendors

Packages can target these AI vendors:

Supported package types

Supported platforms

The Codex and Claude plugins each manage only their own vendor's packages. VS Code and Cursor use the same extension.

DeepSeek Harness packages

The integration targets DeepSeek Harness 0.1.5-rc.3. Select a named profile in aiMarketplace.deepseekHarnessProfile in VS Code. The extension finds a matching local npx installation automatically; aiMarketplace.deepseekHarnessCliPath can point to dsh or its lib/bin.js when installed elsewhere. Set DSH_HOME before launching VS Code if Harness uses a custom home.

Harness skills use SKILL.md under workspace .dsh/skills/<id>/ or $DSH_HOME/skills/<id>/. Rules use RULE.md under .dsh/rules/<id>/ and a Marketplace-owned profile adapter that reads only active rules assigned to that profile. It leaves AGENTS.md untouched. Commands, MCP, agents, and hooks use global delivery through dsh plugin --profile <name> add; each package must ship a prebuilt Harness package.json, cordis.patch.yml, and its referenced modules. The installer does not translate another vendor's format. Harness cloud delivery is unsupported.

For a native Harness page, build with npm.cmd run build:harness, then add the Harness bundle to the selected profile. The Marketplace page appears in the Web UI right Sidebar guide. Repository sources for that page are configured in ~/.ai_marketplace/deepseek-harness.json; credentials come from the same provider environment variables as the Codex and Claude CLIs.

Visual Studio IDE

The Visual Studio extension supports Community, Professional, and Enterprise editions of Visual Studio on Windows x64 and Arm64. It uses the open solution or Open Folder directory for workspace operations; with no workspace, global operations remain available. Repository credentials are stored in Windows Credential Manager and are passed to the isolated sidecar only through authenticated host callbacks.

Repository layout

packages/
  marketplace-core/       Shared catalog, validation, planning, and lifecycle engine
  marketplace-node-cli/   Shared CLI, storage, locking, and host adapters
  vscode-plugin/          VS Code and Cursor extension
  visualstudio-plugin/    Visual Studio 2022+ VSIX
plugins/
  ai-marketplace/         Codex plugin
  ai-marketplace-claude/  Claude Code plugin
  ai-marketplace-harness/ DeepSeek Harness bundle
schemas/                  Generated package-manifest JSON Schema

All interfaces coordinate package state through .ai_marketplace/installed.json and .ai_marketplace/.operation.lock.

Setup and installation

Prerequisites

• Node.js 22 or newer
• A supported Git provider: GitHub, Azure DevOps, or GitLab
• A provider credential only when browsing private repositories, supplied through your host's secret storage or environment

Develop locally

npm.cmd install
npm.cmd test
npm.cmd run lint
npm.cmd run compile

Use VS Code's Run Extension launch configuration to debug the VS Code extension. Do not run out/extension.js directly.

Install in VS Code or Cursor

After the first public release:

• VS Code: install Idov31.ai-marketplace from the Visual Studio Marketplace.
• Cursor: install the same extension from Open VSX through Cursor's Extensions view.

VS Code and Cursor use their native extension update mechanisms. For local development, build dist/ai-marketplace-<version>.vsix with npm.cmd run package:vscode and use Extensions: Install from VSIX....

Install in Visual Studio

On Windows, run powershell -File scripts/download-visualstudio-runtimes.ps1 once, then npm.cmd run package:visualstudio. Close Visual Studio and open dist/ai-marketplace-visualstudio-<version>.vsix to install it. Open the tab using Tools > Open AI Marketplace. Settings are available through Tools > AI Marketplace Settings... and Tools > Options > AI Marketplace.

Visual Studio starts with no repository configured and does not use a hardcoded fallback or install repository-designated default packages. Open the marketplace's Auto Updates tab and find Marketplace sources. Enter a source ID, repository URL, and branch, then select Add a repository and confirm the plan. Multiple GitHub, Azure DevOps, and GitLab sources are supported. Repository configuration is saved in Visual Studio settings; credentials remain in Windows Credential Manager.

Use Set repository credential... in the tool window's native toolbar (or Tools > Set Repository Credential...) to save a shared provider credential or a credential for a configured repository. Supported methods match VS Code: GitHub PAT/access token, Azure DevOps PAT or Microsoft Entra/OAuth access token, and GitLab access token or OAuth access token. Token input is masked, and the dialog can remove saved credentials. Saving a credential from the toolbar refreshes the catalog and reconnects the page. Repository-specific credentials are used after shared authentication is rejected.

The tool window recreates its browser after docking, changing native parents, or reopening. Reload marketplace remains available outside the browser if the page or browser process fails. Reconnecting resets page-local selections and unconfirmed review dialogs; saved repository settings and installed packages are preserved.

The development artifact is unsigned. Publisher signing and public Visual Studio Marketplace publication are intentionally deferred. A WebView2 Runtime is required by Visual Studio.

Install the Codex plugin

After its listing is approved, install AI Marketplace from the OpenAI Plugins Directory in Codex. Public directory releases are reviewed and published through the OpenAI Platform.

For repository development builds only:

codex plugin marketplace add Idov31/ai-marketplace --ref master
codex plugin add ai-marketplace@ai-marketplace-dev

Then ask Codex to Open AI Marketplace. The plugin starts a workspace-local dashboard in Codex's in-app browser. See the Codex plugin guide for updates and CLI details.

Install the Claude Code plugin

After its community listing is approved, install it with:

/plugin marketplace add anthropics/claude-plugins-community
/plugin install ai-marketplace@claude-community

For repository development builds, use:

/plugin marketplace add Idov31/ai-marketplace
/plugin install ai-marketplace@ai-marketplace
/ai-marketplace:manage-packages

See the Claude Code plugin guide for its CLI workflow.

Configure package repositories

New installs start with no configured package repository. Add one from the VS Code Marketplace Configuration tab or configure aiMarketplace.repositories directly.

In VS Code or Cursor, configure aiMarketplace.repositories:

{
  "aiMarketplace.repositories": [
    {
      "id": "company-packages",
      "url": "https://github.com/example/ai-packages",
      "branch": "main"
    },
    {
      "id": "azure-packages",
      "url": "https://dev.azure.com/example/research/_git/ai-packages"
    },
    {
      "id": "private-gitlab",
      "provider": "gitlab",
      "url": "https://gitlab.example.com:8443/research/ai-packages"
    }
  ],
  "aiMarketplace.defaultPlatform": "codex"
}

Each source needs a unique id and a GitHub, Azure DevOps, or GitLab URL. GitHub also accepts owner/repo. For private sources, use AI Marketplace: Set Repository Credential in VS Code/Cursor. The Codex and Claude plugins read provider credentials from environment variables such as GH_TOKEN, AZURE_DEVOPS_ACCESS_TOKEN, and GITLAB_TOKEN.

The VS Code Marketplace Configuration tab can add, edit, and remove these repositories. It saves the list in User Settings, removes older workspace overrides on save, and treats an explicitly empty list as no catalog sources. The same tab stores package auto-update and automatic-group preferences as application-scoped User Settings; legacy values in .ai_marketplace/installed.json are discarded.

Package manifest

Every package folder must contain ai_marketplace.yaml. Only this canonical, grouped manifest format is supported.

schema_version: 1
minimum_reader_schema_version: 1

package:
  name: "@research/example-skill"
  type: skill
  version: "1.0.0"
  description: Reusable guidance for an AI workflow.
  entrypoint: SKILL.md

targets:
  platforms: [codex, cursor, github-copilot, claude]
  delivery: [workspace, global]

metadata:
  tags: [research, workflow]

See the generated manifest schema for every supported field. Package contents are untrusted: paths, YAML structure, and manifest values are validated before installation.

Build packages

npm.cmd run package:vscode
npm.cmd run package:codex
npm.cmd run package:claude
npm.cmd run package:visualstudio
npm.cmd run package:harness
npm.cmd run package

The last command creates all distributable artifacts.