Back to Eliza

Choosing the Right Extension Point

packages/docs/plugins/decision-guide.md

2.0.16.1 KB
Original Source

elizaOS provides multiple ways to extend agent behavior. This guide helps you pick the right one.

Quick Decision Tree

"I want my agent to DO something when asked"Action

"I want to inject context into every response"Provider

"I need a background process running"Service

"I want to add knowledge/instructions without code"Skill

"I need an HTTP endpoint"Route

"I want to react to system events"Event Handler

"I want to assess response quality"Evaluator


Comparison Table

FeatureActionProviderServiceSkillRoute
Triggered byUser message (LLM selects)Every inference cyclePlugin initUser message (LLM selects)HTTP request
ReturnsActionResultContext string--Agent responseHTTP response
Has lifecycleNoNoYes (start/stop)NoNo
Needs TypeScriptYesYesYesNo (markdown)Yes
Hot-reloadableRebuild + restartRebuild + restartRebuild + restartEdit markdown + restartRebuild + restart
Runs in backgroundNoNoYesNoNo

Actions

Use when the agent should perform a task in response to user input. The LLM selects actions from registered options based on description and examples.

typescript
import type { Action } from '@elizaos/core';

const sendEmailAction: Action = {
  name: 'SEND_EMAIL',
  description: 'Send an email to a specified recipient',
  similes: ['EMAIL', 'MAIL', 'MESSAGE'],
  validate: async (runtime, message) => {
    return !!runtime.getSetting('SMTP_HOST');
  },
  handler: async (runtime, message, state) => {
    // Parse recipient and body from message, send email
    return { success: true, text: 'Email sent!' };
  },
};

Good for: API calls, data mutations, tool use, file operations, external service integration


Providers

Use when you need to inject information into the agent's context before every response. Providers run automatically on each inference cycle.

typescript
import type { Provider } from '@elizaos/core';

const timeProvider: Provider = {
  name: 'current-time',
  description: 'Provides current date and time',
  position: -10,
  get: async (runtime, message) => ({
    text: `Current time: ${new Date().toISOString()}`,
  }),
};

Good for: Real-time data, user preferences, system status, database lookups, environment context


Services

Use when you need a long-running background process with startup and shutdown lifecycle.

typescript
import { defineService } from '@elizaos/core';

const webhookService = defineService({
  serviceType: 'webhook-listener',
  description: 'Listens for incoming webhooks',
  start: async (runtime) => {
    // Start HTTP listener, WebSocket connection, etc.
  },
  stop: async () => {
    // Clean up connections and resources
  },
});

Good for: WebSocket connections, polling, cron jobs, queue consumers, cache management


Skills

Use when you want to extend agent behavior with instructions rather than executable code. Skills are markdown-based and don't require TypeScript.

markdown
---
name: git-helper
description: Help users with git commands and workflows
---

When asked about git, provide clear explanations and commands.
Always suggest safe operations first (status, log, diff before reset, force-push).

Good for: Domain knowledge, workflows, instruction sets, prompt engineering, task procedures


Routes

Use when you need to expose HTTP endpoints from your plugin.

typescript
import type { Route } from '@elizaos/core';

const healthRoute: Route = {
  type: 'GET',
  path: '/my-plugin/health',
  public: true,
  handler: async (req, res) => {
    res.json({ status: 'ok', uptime: process.uptime() });
  },
};

Good for: Webhooks, status pages, plugin APIs, file serving, external integrations


Event Handlers

Use when you need to react to system events (messages, connections, actions).

typescript
import type { Plugin } from '@elizaos/core';

const analyticsPlugin: Plugin = {
  name: 'analytics',
  events: {
    MESSAGE_RECEIVED: [
      async (payload) => {
        const { runtime, message, source } = payload;
        // Log message analytics
      },
    ],
    ACTION_STARTED: [
      async (payload) => {
        // Track action usage
      },
    ],
  },
};

Available events: MESSAGE_RECEIVED, VOICE_MESSAGE_RECEIVED, WORLD_CONNECTED, WORLD_JOINED, ACTION_STARTED, ACTION_COMPLETED

Good for: Logging, analytics, side effects, notifications, audit trails


Evaluators

Use when you need to assess response quality or trigger follow-up actions after the agent responds.

typescript
import type { Evaluator } from '@elizaos/core';

const sentimentEvaluator: Evaluator = {
  name: 'sentiment-check',
  description: 'Assess sentiment of agent responses',
  alwaysRun: true,
  validate: async (runtime, message) => true,
  handler: async (runtime, message) => {
    // Analyze response sentiment, log metrics, trigger alerts
  },
};

Good for: Quality monitoring, compliance checks, learning signals, post-response side effects


Combining Extension Points

Many plugins use multiple extension points together:

Plugin TypeTypical Combination
API IntegrationAction (API calls) + Provider (status context) + Service (token refresh)
Platform ConnectorService (connection lifecycle) + Event Handler (messages) + Route (webhooks)
MonitoringEvaluator (quality checks) + Provider (metrics context) + Route (dashboard)
KnowledgeProvider (context injection) + Skill (instructions)