SDK Configuration
Step 3: Select your AI provider and model.
Agent Workshop supports several AI providers. Each has different strengths, pricing, and model options.
Claude Agent SDK (Recommended)
The Claude Agent SDK is Anthropic's official framework for building AI agents. It's the recommended choice for most use cases.
Available Models
| Model | Context | Pricing (Input/Output) | Best For |
|---|---|---|---|
| Claude Sonnet 4.5 | 200K tokens | $3 / $15 per million | General use, best balance of cost and capability |
| Claude Haiku 4.5 | 200K tokens | $1 / $5 per million | Fast, cost-effective tasks |
| Claude Opus 4.1 | 200K tokens | $15 / $75 per million | Complex reasoning, highest capability |
Advantages
- Native streaming support for real-time responses
- Built-in file operation tools
- Optimized for agentic workflows
- Simple setup process
Setup Requirements
- Create an account at console.anthropic.com
- Generate an API key
- Set
ANTHROPIC_API_KEYin your.envfile
OpenAI Agents SDK
The OpenAI Agents SDK provides agent capabilities with function calling and tool use.
Available Models
| Model | Context | Best For |
|---|---|---|
| GPT-5.1 | 128K tokens | Latest capabilities, general use |
| GPT-5 mini | 128K tokens | Cost-effective, faster responses |
| GPT-4.1 | 128K tokens | Stable, well-tested |
Advantages
- Official OpenAI agent framework
- Advanced function calling
- Streaming support
- Wide ecosystem compatibility
Setup Requirements
- Create an account at platform.openai.com
- Generate an API key
- Set
OPENAI_API_KEYin your.envfile
GitHub Copilot SDK
The @github/copilot-sdk package drives the same agent runtime as Copilot CLI. It is the only provider that needs no API key — it authenticates with your existing GitHub Copilot subscription.
Available Models
| Model | Context | Best For |
|---|---|---|
| Auto | Model dependent | Recommended — the runtime picks the best available model |
| Claude Sonnet 4.5 | 200K tokens | Balanced coding and agentic work |
| GPT-5 / GPT-5.4 | 128K tokens | Frontier reasoning, complex multi-step tasks |
| GPT-5.2 Codex | 128K tokens | Long-running coding and refactoring |
The exact model list depends on your subscription. Call client.listModels()at runtime for the authoritative list.
Advantages
- No API key — uses your GitHub Copilot subscription
- Built-in file, search and shell tools provided by the runtime
- Automatically reads
.github/copilot-instructions.md,AGENTS.mdandCLAUDE.md - Bundles the Copilot CLI — no separate runtime install
Setup Requirements
- An active GitHub Copilot subscription
- Node.js 20.19+ (or 22.12+)
- Sign in once with
npx @github/copilot, or set a GitHub token with Copilot access (GITHUB_TOKEN,GH_TOKEN, orCOPILOT_GITHUB_TOKEN)
Copilot uses built-in tools
Copilot agents do not generate file-operations.ts, command-runner.ts or web-tools.ts. The runtime's own tools do that work. Tools you leave disabled in the wizard are excluded from the session and denied at the permission layer, so your permission policy still applies.
Which should I choose?
Claude, OpenAI and Copilot all produce functionally similar TypeScript agents. Choose based on:
- Your preferred AI provider
- Existing API keys or subscriptions you have
- Specific model capabilities you need
- Pricing considerations — Copilot is billed through your existing subscription
Advanced Settings
Max Tokens
Controls the maximum length of agent responses. Range: 1,000 - 8,000 tokens. Default: 4,096 tokens.
- Lower values (1,000-2,000) - Concise responses, faster, cheaper
- Higher values (6,000-8,000) - Detailed responses, better for complex tasks
Temperature
Controls response randomness. Range: 0.0 - 1.0. Default: 0.7.
- Lower values (0.0-0.3) - Deterministic, consistent, best for code/analysis
- Higher values (0.7-1.0) - Creative, varied, best for content generation
Recommended settings
For most agent use cases, the defaults work well. Consider lowering temperature to 0.3 for code-focused agents where consistency matters.