convex-agents-playground（凸 Agent 游乐场）

Purpose

The Playground is a graphical interface for interacting with agents. It lets you browse threads, send messages, inspect tool calls, and adjust context parameters without writing code.

When to Use This Skill

• Testing agents during development
• Manually verifying agent behavior
• Debugging unexpected responses
• Experimenting with prompts and context options
• Demoing agents to stakeholders
• Inspecting tool calls and results
• Analyzing message structure and metadata

Set Up Playground API Endpoint

Create a Convex function that exposes playground APIs:

// convex/playground.ts
import { definePlaygroundAPI } from "@convex-dev/agent";
import { components } from "./_generated/api";
import { weatherAgent, supportAgent, creativeAgent } from "./agents";

export const {
  isApiKeyValid,
  listAgents,
  listUsers,
  listThreads,
  listMessages,
  createThread,
  generateText,
  fetchPromptContext,
} = definePlaygroundAPI(components.agent, {
  agents: [weatherAgent, supportAgent, creativeAgent],
});

Generate API Key

# Generate API key from CLI
npx convex run --component agent apiKeys:issue '{"name":"my-key"}'

# Output:
# { "id": "abc123def456", "name": "my-key" }

Use Hosted Playground

Access the online playground:

1. Visit: https://get-convex.github.io/agent/
2. Enter your Convex deployment URL (from .env.local)
3. Enter your API key
4. Select your playground path (default: "playground")
5. Choose an agent and user to test

Run Playground Locally

# Install and run local playground
npx @convex-dev/agent-playground

# Uses VITE_CONVEX_URL from .env.local
# Open browser to http://localhost:5173

Playground Features

1. User Selection
   - Pick a user to view their threads
   - See all conversations for that user

2. Thread Browsing
   - List all threads for selected user
   - View thread metadata (title, summary)
   - Click to open a thread

3. Message History
   - View all messages in thread
   - See message ordering and step order
   - Inspect message metadata
   - View tool call details

4. Send Message
   - Type a message to test agent
   - Configure save options
   - Send and see response in real-time
   - View usage information

5. Context Configuration
   - Adjust recentMessages count
   - Enable/disable text search
   - Enable/disable vector search
   - Set search result limits
   - Test different context options

6. Tool Call Analysis
   - View tool calls and parameters
   - See tool results
   - Inspect tool execution details

Test Multiple Agents

Switch between agents in playground:

// convex/playground.ts
export const {
  isApiKeyValid,
  listAgents,
  // ...
} = definePlaygroundAPI(components.agent, {
  agents: [fastAgent, creativeAgent, analyticalAgent, supportAgent],
});

Manage API Keys

Generate and revoke keys:

# Generate new key (or reissue existing)
npx convex run --component agent apiKeys:issue '{"name":"dev-key"}'

# Reusing a name revokes and reissues

# Multiple keys for different environments:
npx convex run --component agent apiKeys:issue '{"name":"prod-key"}'
npx convex run --component agent apiKeys:issue '{"name":"staging-key"}'

Key Principles

• Security via API keys: Keys needed to access playground
• Read-only by default: Doesn't modify data unless you send messages
• Real-time updates: See agent responses immediately
• Context control: Experiment with context settings live
• No code needed: Use without writing any code

Development Workflow

1. Define agent (convex/agents.ts)
2. Export in playground (convex/playground.ts)
3. Generate API key
4. Open playground
5. Test agent
6. Debug behavior
7. Adjust agent
8. Repeat

Next Steps

• See debugging for logging context
• See fundamentals for agent setup
• See rate-limiting for understanding costs