Connect a model API

Connect a model

Connect a model provider to createHarness() so the built-in loop can call the model API. Then compose the harness and a model with createAgent() and pass that agent to a dispatch.

import {
  createAgent,
  createAnthropicModelProvider,
  createHarness,
} from "@elie-laloum/outpost";

export const agent = createAgent({
  model: { name: "claude-sonnet-5-5", maxOutputTokens: 16_000 },
  harness: createHarness({
    modelProvider: createAnthropicModelProvider({
      apiKey: process.env.ANTHROPIC_API_KEY ?? "",
    }),
  }),
});

Anthropic requires maxOutputTokens, hence the object form of model. Each model request is an HTTP call from your Node.js process; tools still run in the sandbox that the dispatch allocates.

Choose a protocol

baseUrl is required for OpenAI and includes the version prefix. The protocol you choose is the only one used: an error never switches to another.

import { createOpenAIModelProvider } from "@elie-laloum/outpost";

const openai = createOpenAIModelProvider({
  baseUrl: "https://api.openai.com/v1",
  api: "responses",
  apiKey: process.env.OPENAI_API_KEY ?? "",
});

Factory	api	Path added to baseUrl	Use it for
createOpenAIModelProvider	"chat-completions" (default)	/chat/completions	OpenAI and servers that expose Chat Completions.
createOpenAIModelProvider	"responses"	/responses	The OpenAI Responses API.
createAnthropicModelProvider	none	/messages	The Anthropic Messages API; baseUrl defaults to https://api.anthropic.com/v1.

createCodexHarness({ modelProvider }) is a different setting: it points the Codex CLI, inside the sandbox, at a Responses-compatible service (Codex).

Use a local endpoint

Set apiKey: false for a server without authentication. The address is resolved from your host, not from the sandbox.

import { createOpenAIModelProvider } from "@elie-laloum/outpost";

const local = createOpenAIModelProvider({
  baseUrl: "http://127.0.0.1:8080/v1",
  apiKey: false,
});

Keep the key on the host

Pass the model API key explicitly to the provider. These providers do not load account logins or read environment variables automatically. The key stays in your Node.js process; Authentication explains how this differs from CLI agents.

Set the model and its reasoning

The agent’s model is a name or { name, reasoning, maxOutputTokens }. createAgent() rejects settings the provider does not support.

API reference: AgentModel.

The service still checks the model name and levels on each request.

Stream text as it arrives

Both providers stream. The harness emits text-delta events while the model writes, and a reasoning event when a response contains readable reasoning. Print them from observe with if (event.kind === "text-delta") process.stdout.write(event.text) (Follow progress).

Bound each request

API reference: OpenAIModelProviderOptions and AnthropicModelProviderOptions.

A timeout fails with code timeout. The harness streams with both providers, so a long answer that keeps arriving never times out: bound the whole turn with limits.

Cache the prompt prefix

The harness’s cache option, on by default, asks the provider to reuse the conversation prefix between steps.

• Anthropiccache marks the request for caching; cacheSystem: true adds a breakpoint on the harness instructions, which must then exist.
• OpenAIOutpost sends no cache field; OpenAI caches stable prefixes on its side.
• UsageCache reads appear in usage.cached, and Anthropic cache writes in usage.cacheCreated.

A cache hit is never guaranteed.

Retry after rate limits and outages

A provider sends each request once. When the service answers HTTP 429 with Retry-After, the error keeps that delay: a task retry waits at least that long, and a quota pause resumes at that time.

Rate limits fail with code quota; overloads, 5xx and connection failures are marked unavailable for fallback agents. Quota pauses list what counts as a quota.

Connect another API

Implement ModelProvider: request() returns a result, stream() and validate() are optional.

Limits

• Reasoning is replayed only to the same provider, endpoint and model. Changing one drops it from the history.
• baseUrl cannot contain credentials, a query or a fragment. Redirects are refused.
• Anthropic responses with content other than text, tool calls and thinking fail with response.

API: createOpenAIModelProvider · createAnthropicModelProvider · OpenAIModelProviderOptions · AnthropicModelProviderOptions · ModelProvider · AgentModel.