Write a plugin

The two are easy to confuse, and picking the wrong one wastes a lot of work. The difference fits in one sentence: if it takes code, write a plugin; if explaining it clearly is enough, write a Skill.

A Skill is plain text (a prompt plus trigger words): it goes live on submission with no review, and is much faster to write. If a Skill covers your need, go write a Skill and skip the rest of this page.

A plugin is a Java project compiled into a JAR, plus a manifest.json, zipped up and submitted. You need JDK 21 and Maven.

Download the template project above — it is itself a complete working example. Only two files matter. The first is the tool class:

package com.example.myplugin;

import dev.langchain4j.agent.tool.Tool;

/**
 * Plugin tool class.
 *
 * Three hard rules — break any one of them and the tool will not be visible to the AI:
 * 1. Must have a no-arg constructor (instantiated via reflection by the host);
 * 2. Annotate tool methods with @Tool; the method name is the tool name and must match manifest.json tools[].name;
 * 3. Use String parameters and return values for simplicity (serialize complex structures as JSON strings).
 *
 * The @Tool description is read by the AI and determines when it invokes this tool.
 * State clearly when to call it rather than only how it is implemented.
 */
public class MyTools {

    @Tool("Count characters and CJK characters in a text passage and return a readable summary. Call when the user asks for character or word counts.")
    public String countChinese(String text) {
        if (text == null || text.isBlank()) {
            return "Empty input: 0 characters";
        }
        long cjk = text.codePoints()
                .filter(cp -> Character.UnicodeScript.of(cp) == Character.UnicodeScript.HAN)
                .count();
        return String.format("Total %d chars, including %d CJK chars", text.length(), cjk);
    }
}

La description dans @Tool est écrite pour l'IA — elle détermine directement si l'IA appelle votre outil au bon moment. Écrire « quand l'utiliser » est bien plus utile que « ce que fait cette méthode » : c'est le premier élément que l'IA doit évaluer.

Le second est manifest.json, qui décrit ce qu'est le plugin :

{
  "id": "my-plugin",
  "name": "My Plugin",
  "version": "1.0.0",
  "description": "One sentence explaining what this plugin does for users, shown on the marketplace card.",
  "author": "Your Name or Team",
  "homepage": "https://example.com",
  "permissions": [],
  "tools": [
    {
      "name": "countChinese",
      "description": "Count characters in text",
      "permissions": []
    }
  ],
  "backendJars": ["my-plugin-1.0.0.jar"]
}

Les noms des outils doivent correspondre des deux côtés : tools[].name doit être identique au nom de la méthode Java. S'ils diffèrent, l'outil ne s'enregistre pas et l'IA ne peut pas le voir.

Ensuite, empaquetez-le :

mvn package
mkdir -p dist && cp target/my-plugin-1.0.0.jar manifest.json dist/
cd dist && zip -r ../my-plugin-1.0.0.zip . && cd ..

Notez que le zip doit contenir les fichiers eux-mêmes — n'imbriquez pas de répertoire supplémentaire. La décompression doit révéler manifest.json directement, et non un dossier.

Pas besoin d'attendre la validation. Copiez l'intégralité du répertoire dist/ dans votre dossier local de plugins et redémarrez AI WorkDeck :

# macOS / Linux
~/.aiworkdeck/plugins/my-plugin/

Votre plugin devrait apparaître sous Place de marché des plugins → Installés. Activez-le, puis posez une question dans une conversation qui devrait utiliser votre outil et vérifiez si l'IA l'appelle. Si ce n'est pas le cas, la description @Tool n'explique probablement pas correctement quand l'utiliser.

Quatre valeurs facultatives ; déclarez ce que vous utilisez :

• file_readLire les fichiers du projet
• file_writeCréer, modifier ou supprimer des fichiers
• networkAccéder au réseau externe
• editorAgir sur l'éditeur de documents

Chaque version est examinée par un humain, généralement en un à deux jours ouvrés. Une analyse automatisée est d'abord exécutée, et son rapport se trouve à côté de votre déclaration de permissions sur le bureau de l'examinateur.

These are rejected outright:

• Declared permissions don’t match the APIs actually called
• Custom TrustManager or any other way of bypassing certificate validation
• Hardcoded IP addresses, or data sent out over plaintext HTTP
• Reflection into the host’s internal objects (database connections, config services, etc.)
• Obfuscated or packed code, or anything that hides what the code does
• Version number not higher than the previous one

Once approved, the platform signs the whole package with its private key and clients verify the signature on install. After that, nobody — including us — can alter the package contents without re-signing.

If a problem surfaces after release, we revoke that version. Clients pick up the revocation list, disable it automatically and notify the user.

Our users are lawyers, and their machines hold clients' confidential material. A plugin gets the same access as the host app — far beyond what it needs for itself.

So review errs on the strict side, and rejections come with reasons. If your plugin genuinely needs a sensitive-looking capability, explain why in the submission notes — it makes review much faster.

If what you want is an interface — a form, a lookup tool, a dashboard — you do not need Java. Put a web/ directory in the package (plain static HTML/JS/CSS, entry web/index.html) and point manifest frontendEntry at it. It may ship no JAR at all.

Here the permissions are real. The desktop app loads a Web plugin into a sandboxed iframe, deliberately not same-origin: plugin scripts never see the host session, every capability travels over the postMessage bridge, and the host checks each call against the manifest permissions. Unlike a JAR plugin, this declaration is an enforcement boundary rather than a self-description.

web/awd-plugin-sdk.js in the template wraps that bridge. Nothing works before the handshake:

<script src="awd-plugin-sdk.js"></script>
<script>
  const ctx = await awd.ready();                   // { pluginId, projectId, language, theme, themeTokens }
  const files = await awd.files.list();            // requires file_read
  const doc = await awd.files.read(files[0].path); // { path, content, truncated }
  await awd.ui.toast('Completed');
  await awd.storage.set('draft', { title: 'x' });  // plugin-scoped KV, capped at 64 KB
</script>

Errors arrive as rejected promises with err.code: permission_denied when the manifest does not declare what the call needs, unknown_method when the host does not recognize it. files.read caps text at 5 MB (truncated becomes true), and storage caps each plugin at 64 KB total.

Theme channel (v2.6)

The handshake context now carries theme ('light'|'dark') and themeTokens (a table of --awd-* CSS variables); the host also pushes { type: 'theme', theme, tokens } on every later switch. The SDK applies it automatically — setting data-theme, toggling the awd-theme-light/awd-theme-dark class, and writing each token as a CSS custom property — so plugin CSS can just use var(--awd-surface) with zero JS. For script-driven behavior, use awd.theme.get() to read the current value and awd.theme.onChange(cb) to subscribe. The host simulator has a light/dark toggle button in its header, so you can test both themes without installing the desktop app.

The template also ships a host simulator, dev/host-simulator.html: it impersonates the desktop host, performs the handshake, serves every method with fake data, and logs each bridge message. Run a local static server and you can develop in a browser without installing AI WorkDeck.

python3 -m http.server 8000
# -> http://localhost:8000/dev/host-simulator.html

Opening it directly over file:// usually shows a blank frame — browsers refuse to load sandboxed child documents from file://. The server must not rewrite URLs either: npx serve enables clean URLs by default, redirecting /web/index.html to /web so the relative SDK reference 404s.

The Web download includes a standalone Node.js 20+ CLI. Keep scripts/ and sdk/plugin-sdk/ in their relative locations. Test UI in the simulator, then permissions, project switching, disabling, cancellation and reconnection in a compatible isolated host. Example domains are placeholders, not working services.

Declare sidebar/workspace views with project/global scope; one view per placement and an existing web/ entry. Sidebar docks are left/right/bottom. Settings default to installation scope; user/project settings may be changed by the plugin. Installation settings remain in trusted host UI. Secrets never return to the iframe; use connections for per-user credentials.

Connections declare a fixed public HTTPS origin and fixed GET/POST operations. Authentication is none, Bearer or public-client OAuth2 PKCE. connect opens trusted host UI; tokens stay in the host. Never pass arbitrary URLs, authorization headers or host setting keys. BYO services and shared host AI/speech channels are separate choices with their own permissions and usage constraints.

OAuth uses Authorization Code with S256 PKCE. A local host uses its actual port at http://127.0.0.1:<port>/api/plugins/oauth/callback; server deployments must set ai.plugins.oauth-redirect-uri to a trusted HTTPS callback and register it with the provider. Expired credentials without refresh or credentials rejected with 401/403 return needsReconnect and require host-side login again.

MCP currently supports public HTTPS Streamable HTTP only: no stdio, local processes, private addresses or redirects. Declare network, every server and tool; remote names and schemas must match. Input roots are objects. Supported schema keywords are type/title/description, object properties/required/boolean additionalProperties, array items and string enum. Types include boolean/integer/number. No $ref, oneOf, nullable or other keywords. Limits: depth 8, 64 properties per object and 64 tools per server. connectionId references a same-origin connection in this plugin. Legacy credentialKey only references a declared installation secret.

Development installs require explicit administrator service trust in the host settings page. Trust is bound to the full manifest SHA256 and expires when it changes. Packages must not contain .awd-dev, .awd-service-trust or .awd-bundled markers. Trust does not unlock development JARs or local processes.

speech, tasks and clipboard permissions grant only catalogued services.request operations; this is not an arbitrary backend proxy. services.cancel cancels this plugin’s requests. Handle unsupported hosts, missing permissions and failures explicitly; do not silently switch to external or paid services.

Platform billing currently covers plugin purchase entitlement only; submitted prices are one-time purchases. There is no platform per-call collection or revenue-sharing API. Developer-owned subscriptions may declare connections.billing status and checkout. An opened checkout is not a successful payment.

Host billing.status refreshes the current account entitlement and returns refreshed; false means the refresh did not succeed and active may come from a still-valid cache, not a fresh payment confirmation. Checkout waits for the browser-opening result and rejects failures. Test actual host login, token expiry, subscription status, MCP errors, account switching and shared/BYOK routing; a synthetic provider verifies host integration, not live OAuth or payment certification.

Complete the free developer eligibility review and generate a submission token in the developer center. Supply it through the local AWD_PLUGIN_TOKEN environment variable. Never put it in source, Agent prompts or packages. Valid submissions remain pending human review and signing; passing local validation is not marketplace approval.

node scripts/plugin-dev.mjs validate .
node scripts/plugin-dev.mjs package . --out ../my-plugin.zip
# Start the synthetic UI simulator, then open /dev/host-simulator.html
python3 -m http.server 8000
# Explicit development installation: select an isolated imported project/folder first.
# Replace PORT with the isolated host port. AWD_SESSION_ID is supplied locally if required.
# Do not paste credentials into source or an Agent prompt.
node scripts/plugin-dev.mjs install --endpoint http://127.0.0.1:PORT --project-id PROJECT_ID --folder-id FOLDER_ID
# Explicit marketplace submission only; provide AWD_PLUGIN_TOKEN locally.
node scripts/plugin-dev.mjs submit . --endpoint https://www.aiworkdeck.com/api/plugins --price-cents 0
node scripts/plugin-dev.mjs submissions --endpoint https://www.aiworkdeck.com/api/plugins

{
  "id": "my-service-plugin",
  "name": "Service panel",
  "version": "1.0.0",
  "description": "SDK 1.6 host preview example",
  "author": "Example developer",
  "frontendEntry": "web/index.html",
  "permissions": [
    "network",
    "ai",
    "speech",
    "tasks",
    "clipboard"
  ],
  "settings": [
    {
      "key": "language",
      "label": "Language",
      "type": "string",
      "scope": "user",
      "default": "zh-CN"
    },
    {
      "key": "summaryStyle",
      "label": "Summary style",
      "type": "string",
      "scope": "project",
      "default": "brief"
    }
  ],
  "contributes": {
    "views": [
      {
        "id": "service-panel",
        "name": "Service panel",
        "entry": "web/index.html",
        "placement": "sidebar",
        "scope": "project",
        "defaultDock": "left",
        "allowedDocks": [
          "left",
          "right",
          "bottom"
        ]
      }
    ],
    "connections": [
      {
        "id": "service",
        "label": "My service",
        "baseUrl": "https://api.example.com",
        "auth": {
          "type": "bearer"
        },
        "operations": {
          "summarize": {
            "method": "POST",
            "path": "/v1/summarize"
          },
          "status": {
            "method": "GET",
            "path": "/v1/subscription"
          }
        },
        "billing": {
          "statusOperation": "status",
          "checkoutUrl": "https://example.com/subscribe"
        }
      }
    ],
    "mcpServers": [
      {
        "id": "remote",
        "transport": "streamable-http",
        "url": "https://api.example.com/mcp",
        "connectionId": "service",
        "tools": [
          {
            "name": "summarize",
            "description": "Summarize supplied text",
            "inputSchema": {
              "type": "object",
              "properties": {
                "text": {
                  "type": "string"
                }
              },
              "required": [
                "text"
              ],
              "additionalProperties": false
            }
          }
        ]
      }
    ]
  }
}

Complete task to give your Agent

Implement this requirement using the downloaded AI WorkDeck Web plugin template: [feature, inputs, outputs, acceptance criteria].
Read README.md, sdk/plugin-sdk/awd-plugin-sdk.d.ts and manifest.schema.json first. Use minimum permissions. SDK 1.6 platform features require the unreleased new host; simulator success does not prove installed-client support.
Edit only this plugin. Preserve scripts/ and sdk/plugin-sdk/ layout. Do not modify the SDK to bypass checks. Support --awd-* theme tokens, keyboard interaction, narrow panels and light/dark themes.
Declare views and installation/user/project settings as needed. Credentials must use trusted host connections; never put tokens in the iframe, logs, repository or zip.
Choose shared host AI/speech/tasks/clipboard with the required permissions, or fixed HTTPS BYO operations with network permission. MCP must use public HTTPS Streamable HTTP and explicitly declared tools with the supported schema subset. No stdio, private hosts, arbitrary URLs or system settings.
Use synthetic simulator data to verify success, permission denial, disconnect, retry and cancel. Do not access customer data, real login or paid services. Use only my designated isolated project for host installation; ask the host administrator to grant service trust in settings after manifest changes.
Run node scripts/plugin-dev.mjs validate . and node scripts/plugin-dev.mjs package . --out ../my-plugin.zip. Verify no .env, credentials, private keys or host trust markers. Report changes, checks, unverified boundaries and zip location.
Only submit when I explicitly request it, using a locally provided AWD_PLUGIN_TOKEN without printing it. Report pending review accurately; never claim approval, payment or release.

MCP Streamable HTTP specification