Run Claude Agent

Octopus.Claude exported 2026-07-22 by Octopus Deploy belongs to ‘Claude’ category.

Runs a Claude agent with a prompt using the Anthropic API.

Parameters

When this step is included in a project’s deployment process, the parameters below can be set.

Prompt

Octopus.Action.Claude.Prompt

The prompt to send to the Claude agent.

API Token

Octopus.Action.Claude.ApiToken

The Anthropic API token for Claude. Stored as a sensitive value.

Model

Octopus.Action.Claude.Model

The Claude model to use. Leave blank to use the Claude CLI’s current default model.

Additional MCP Servers

Octopus.Action.Claude.McpServers

JSON array of MCP server configurations. Each entry has name, type, command, args, and env fields.

Octopus MCP API Key

Octopus.Action.Claude.OctopusMcpApiKey

The Octopus API key the built-in Octopus MCP server uses to access this Octopus instance. Stored as a sensitive value.

Octopus MCP Tools

Octopus.Action.Claude.OctopusMcpTools

JSON array of tool names/patterns from the Octopus MCP server the agent may use. [”*”] allows all tools. Each pattern becomes mcp__octopus__ in the agent’s permissions.

Max Turns

Octopus.Action.Claude.MaxTurns

Optional. Maximum number of agentic turns (default: 10).

Max Budget (USD)

Octopus.Action.Claude.MaxBudgetUsd

Optional. Maximum budget in USD for the agent run.

Permission Mode

Octopus.Action.Claude.PermissionMode

Optional. Controls which permission strategy Claude Code uses. dontAsk: the standard non-interactive behaviour. Auto: classifier-gated auto mode, which also consults the Auto Mode Config below. Leave blank to use dontAsk.

Permissions

Octopus.Action.Claude.Permissions

Optional. Raw JSON for Claude’s permissions block, e.g. {“allow”:[“Read”,“Bash(npm run test:*)”],“deny”:[“WebFetch”,“Read(./.env)”]}. Embedded verbatim into the generated Claude settings file. MCP server tool patterns are added automatically and should not be included here.

Auto Mode Config

Octopus.Action.Claude.AutoModeConfig

Optional. Raw JSON for Claude’s auto mode classifier config (keys such as environment, allow, soft_deny, hard_deny; supports the literal “$defaults” entry), e.g. {“environment”:[“$defaults”,“github.com/acme”],“soft_deny”:[“$defaults”]}. Only consulted when Permission Mode is Auto.

Effort

Octopus.Action.Claude.Effort

Optional. Reasoning effort level: low, medium, high, xhigh, or max.

Sandboxing

Octopus.Action.Claude.SandboxMode

Controls how the Claude Code process is sandboxed. None: no sandboxing, Claude runs with the same permissions as the Tentacle process. Bash: enables Claude Code’s built-in bash sandbox, writing the sandbox settings to .claude/settings.sandbox.json in the working directory to restrict what the bash tool can do. SandboxRuntime: runs Claude inside the Anthropic sandbox-runtime (srt) process for stronger isolation.

Sandbox Settings

Octopus.Action.Claude.SandboxSettings

The sandbox settings as a single JSON document, written verbatim to the settings file for the selected sandbox mode. Only used when Sandboxing is Bash or SandboxRuntime. For Bash this is the Claude Code settings.json content (a top-level “sandbox” object with “enabled”, “network” and “filesystem”); for SandboxRuntime it is the srt settings (top-level “network” and “filesystem”). The UI supplies smart defaults.

Page updated on Wednesday, July 22, 2026