Amp Agent

The ampcode plugin runs Amp in execute mode. Prompts go through stdin rather than command-line arguments, and Claude Code-compatible JSONL output provides text, tool display, and subagent tracing.

WARNING

Amp runs tools without approval by default. Setting dangerouslyAllowAll = false only omits a legacy flag; it does not restrict current Amp's tool access. For untrusted repositories or inputs, use an isolated environment or an Amp policy plugin. See Amp's tools and permissions documentation.

Prerequisites

Install and authenticate Amp using the official instructions. The npm package is @ampcode/cli (@sourcegraph/amp is the legacy package name).

Bash

npm install -g @ampcode/cli
amp --version
amp login

Check the installed version's interface with amp --help before choosing a mode. Authentication is managed by Amp; Ralph TUI does not create accounts or install permission policies.

Usage

Bash

ralph-tui run --prd ./prd.json --agent ampcode

TOML

# .ralph-tui/config.toml
agent = "ampcode"
 
[agentOptions]
mode = "medium"
streamJson = true
dangerouslyAllowAll = false

Option	Type	Default	Description
`mode`	string	Amp's default	Passed to `--mode`; supports built-in and custom modes from your installed CLI
`streamJson`	boolean	`true`	Parse structured output; tracing also enables JSONL regardless of this setting
`dangerouslyAllowAll`	boolean	`false`	Opt in to the legacy `--dangerously-allow-all` flag; false is not a current-Amp permission boundary
`timeout`	number	`0`	Execution timeout in milliseconds; zero means no timeout
`command`	string	`"amp"`	Override the executable path

Current built-in modes include low, medium, high, ultra, and claude-code; custom modes and older CLI mode names are passed through unchanged. Leave mode unset to follow Amp's default. Use agentOptions.mode, not Ralph TUI's --model, because Amp selects models through modes.

If using AMP_API_KEY rather than saved login credentials, explicitly allow it through Ralph TUI's environment filter in the agent configuration:

TOML

[[agents]]
name = "amp"
plugin = "ampcode"
default = true
envPassthrough = ["AMP_API_KEY"]

Output and Completion

Structured mode displays assistant text and tool calls, buffers partial JSONL lines, and forwards parsed events for tracing. Only root assistant text and successful result text count toward <promise>COMPLETE</promise> detection; tool inputs, tool results, thinking, child-agent output, metadata, and malformed protocol records do not complete tasks. An error result fails the execution even if Amp exits with code zero.

Explicit streamJson = false preserves unstructured text output when tracing is off. That mode cannot distinguish assistant text from tool output; use the default structured mode for reliable completion detection.