Configure Claude Code

Install

The agent image contains Claude Code. Remote sandboxes install claude if it is missing before the first turn; local host execution uses the executable already on your PATH.

• Pinned versionThe image and remote installation use @anthropic-ai/claude-code at the version in agentVersions.claude.
• Remote installationnpm installs it under ~/.outpost-tools in the sandbox.
• Image onlybootstrap: false on dispatch() or createSandbox() skips installation; the image must provide claude.

Sign in with your account

Run claude, then /login, on the host. authentication: "account" copies the subscription login from ~/.claude/.credentials.json (or $CLAUDE_CONFIG_DIR/.credentials.json) into the sandbox’s private home.

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

const coder = createAgent({
  harness: createClaudeHarness({ authentication: "account" }),
});

On macOS, Claude Code keeps its login in the keychain by default, and Outpost never reads a keychain. Run claude setup-token instead and pass the printed token as a variable. It uses your Pro, Max, Team or Enterprise plan.

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

const coder = createAgent({
  harness: createClaudeHarness({
    authentication: { account: { variable: "CLAUDE_CODE_OAUTH_TOKEN" } },
    variables: {
      CLAUDE_CODE_OAUTH_TOKEN: process.env.CLAUDE_CODE_OAUTH_TOKEN ?? "",
    },
  }),
});

Choosing between account and API access, and where credentials go: Authentication. Vendor requirements: Claude Code authentication.

Use an API key

authentication: "usage" passes ANTHROPIC_API_KEY. Requests are billed to your Anthropic API account, not to a subscription.

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

const coder = createAgent({
  harness: createClaudeHarness({
    authentication: "usage",
    variables: { ANTHROPIC_API_KEY: process.env.ANTHROPIC_API_KEY ?? "" },
  }),
});

Choose a model and options

Omit model to use Claude Code’s default. A model object adds an effort level and an output limit.

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

const coder = createAgent({
  harness: createClaudeHarness({
    authentication: "account",
    partialMessages: true,
  }),
  model: { name: "sonnet", reasoning: "high", maxOutputTokens: 32_000 },
});

API reference: ClaudeSettings and AgentModel.

Available features

Choose an agent compares these capabilities across agents.

• ConversationsCaptured after each turn, then resumed or forked in a new sandbox. resumeforkresponse repair
• SteeringInstructions join the running turn through its stream-json input, on every sandbox provider.
• Quota pausesA usage limit fails with code quota, with the reset time when Claude reports it.
• Fallback agentsAPI 5xx, overload and connection failures are classified as unavailable.
• ProgressTool calls, thinking and token usage per message, including cache tokens.
• MCP serversStdio and HTTP servers, excluded tools and host OAuth logins.

Limits

• One credential: Declaring ANTHROPIC_API_KEY with account access, or CLAUDE_CODE_OAUTH_TOKEN with API access, fails before the turn starts.
• Permissions: Without permissions, headless turns run with --dangerously-skip-permissions. The sandbox, not the permission mode, is the isolation boundary.
• Duplicate settings: Set maxOutputTokens or a CLAUDE_CODE_MAX_OUTPUT_TOKENS variable, not both.

API: createClaudeHarness · ClaudeSettings · createClaudeConversations · AgentAuthentication.