Choose an agent

The agents

An agent combines a harness with an optional model selection. The harness runs the agent’s loop: it can use an installed CLI, such as Codex or Claude Code, or Outpost’s built-in loop. Pick the agent first, then configure its access and model.

• Claude CodeAnthropic’s coding CLI, with live steering and the widest model settings. accounttokenAPI key
• CodexOpenAI’s coding CLI, with live steering and custom Responses-compatible providers. accountAPI key
• GitHub Copilot CLIGitHub’s coding CLI, billed to your Copilot plan; it resumes but does not fork. accounttoken
• Kimi CodeMoonshot AI’s coding CLI; with an API key, you must name the model. accountAPI keymodel required with API key
• AntigravityGoogle’s agy CLI; its conversations continue only in their open sandbox. accountAPI key
• Built-in harnessOutpost’s own loop, running your tools over an OpenAI or Anthropic API. API keymodel required

Compose an agent

Pass a harness to createAgent(), then add model if you want to override the agent’s default. Outpost provides a harness constructor for each supported CLI: createClaudeHarness(), createCodexHarness(), createCopilotHarness(), createKimiHarness() and createAntigravityHarness().

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

export const reviewer = createAgent({
  harness: createClaudeHarness({ authentication: "account" }),
  model: { name: "opus", reasoning: "high" },
});

Pass the agent to dispatch() as agent. authentication chooses between your account and an API key; see Authentication.

Select a model

API reference: ModelSpec and AgentModel.

createAgent() rejects model settings the harness cannot apply, including unsupported reasoning levels or maxOutputTokens. This check happens before execution. The model service checks your account’s access when the request runs; the API reference below describes the supported settings.

API reference: ModelSpec and AgentModel.

Compare capabilities

Sign in

Choose the credential you already have. Authentication explains how to pass it to the sandbox and how account access differs from API billing.

Agent	Account login	Account token in a variable	API key
Claude Code	Yes	Yes	Yes
Codex	Yes	No	Yes
Copilot CLI	Yes	Yes	No
Kimi Code	Yes	No	Yes, with an explicit model
Antigravity	Yes	No	Yes
Built-in harness	No	No	Yes

Continue work

Every agent can continue a conversation while its sandbox remains open. To resume in a new sandbox, Outpost also needs to capture and restore that conversation. A fork starts a separate conversation from the same context.

Agent	Resume in a new sandbox	Fork	Steering mode
Claude Code	Yes	Yes	injected
Codex	Yes	Yes	injected
Copilot CLI	Yes	No	resumed
Kimi Code	Yes	Yes	resumed
Antigravity	No	No	resumed
Built-in harness	Yes	Yes	injected

With injected, new instructions reach the running turn. With resumed, Outpost stops the turn and resumes its conversation. See steering to send instructions and conversations to resume or fork.

All these agents support response repairs when conversation continuation is enabled. Setting saveConversations: false on Claude Code or Codex keeps only continuation in the same sandbox. Setting conversations: false on the built-in harness disables continuation, forks and repairs.

For progress events and token counters, see progress reporting and budgets. Quota reset times depend on the agent or model provider; quota pauses explains when a workflow can wait automatically.

Give agents MCP tools

Every harness accepts mcpServers: stdio commands or HTTP endpoints, with secrets passed by variable name. MCP servers shows how to declare them.

For OAuth, Claude Code, Codex and Kimi can use a login saved on the host; the built-in harness uses client credentials. Copilot CLI and Antigravity do not support this OAuth configuration. See MCP server login.

Fall back to another agent

createFallbackAgent([...], { on }) hands a dispatch to the next candidate when a usage limit or an outage stops the current one. See Fallback agents.

Run without a CLI

createHarness() drives a model API directly, with the tools, permissions and subagents you declare. It installs no CLI and requires a model provider. See Built-in harness.

API: createAgent · AgentOptions · ModelSpec · AgentModel · createClaudeHarness · createCodexHarness · createCopilotHarness · createKimiHarness · createAntigravityHarness · createHarness · createFallbackAgent.