@openclaw/ai example: minimal chat completion

Smallest possible consumer of the published @openclaw/ai package: creates an isolated runtime, registers the built-in API providers, and streams one completion. No OpenClaw application code is involved — inside this repo the workspace link resolves to the package's built dist, the same artifact npm consumers install, so pnpm build must have run first.

Choose a provider and replace its placeholder with a real API key.

POSIX shell commands (the Ollama command also works in PowerShell):

ANTHROPIC_API_KEY=example-anthropic-key-not-real node index.mjs "Say hello"
OPENAI_API_KEY=example-openai-key-not-real node index.mjs --provider openai "Say hello"
# keyless, against a local Ollama server (OLLAMA_MODEL overrides the model id)
node index.mjs --provider ollama "Say hello"

PowerShell (assignments remain in this session until removed or the shell closes):

$env:ANTHROPIC_API_KEY = "example-anthropic-key-not-real"
node index.mjs "Say hello"

$env:OPENAI_API_KEY = "example-openai-key-not-real"
node index.mjs --provider openai "Say hello"

# Clear the session-scoped keys.
$env:ANTHROPIC_API_KEY = $null
$env:OPENAI_API_KEY = $null

Text deltas stream to stdout; stop reason and token usage go to stderr.

Notes for library consumers:

• createLlmRuntime() gives you an isolated registry; nothing is registered globally by importing the package.
• registerBuiltInApiProviders(runtime.registry) opts into the eight built-in transports (Anthropic, OpenAI Completions/Responses, Azure OpenAI, ChatGPT Responses, Google, Vertex, Mistral). Provider SDK modules load lazily on first use.
• Host policy (custom fetch, secret redaction, strict-tool defaults, logging) is injectable via configureAiTransportHost; the defaults are inert.
• Full TypeScript types ship with the package; this example uses plain ESM so it runs with bare node.

Source: examples/ai-chat/README.md