Amber Agent - Claude code compatible web CLI

Light+ Snake Demo

I built Amber because I wanted a coding agent that felt like a terminal, but with a browser UI to take advantage of linking. Also Claude Code has a serious amount of bloat that I wanted to cut out.

So I reverse-engineered Claude Code and rebuilt it from scratch as a minimal web CLI, with prompt-accurate reproduction of the workflow.

Amber takes advantage of the fact that it's in a web browser to do many things that CLIs cannot do easily, such as linking forked sessions together with hyperlinks, and linking agent sessions to the main session and vice versa.

What makes it different:

• Amber takes advantage of hyperlinks everywhere.
• Supports OpenAI Codex plans and pretty much any other provider with OpenAI API and Anthropic API support.
• Only 4 runtime npm dependencies — diff, markdown-it, smol-toml, and yaml. No framework, no database, no bloat.
• Easily configurable sub-agents. You can even customize the model per each agent.
• Claude Code, rebuilt. Reverse-engineered from Claude Code's behavior, reproducing the prompt flow accurately while stripping it down to a lightweight tool.
• Supports Claude Code Plugins

Themes

It supports 4 themes out of the box. light+ follows the VS Code Light+ palette; the other alternate themes are shown below.

Theme = Amber

Theme = hacker

Theme = light

Install and run

Prebuilt binary

Linux x86_64 and Apple Silicon macOS are supported.

curl -fsSL https://raw.githubusercontent.com/tpolasek/amber/main/install.sh | sh
export PATH="$HOME/.local/bin:$PATH"
amber

The first run creates ~/.amber/settings.toml and opens Amber's Settings modal. Add your provider credentials and save; Amber validates and loads them without a restart.

Run from source

Requires Node.js 18.17 or newer.

git clone https://github.com/tpolasek/amber.git
cd amber
npm install
npm run build
npm start

For development with automatic rebuilds:

npm run dev

Settings

Open Settings with the gear beside the selected model. The modal is the primary interface for ~/.amber/settings.toml and exposes every supported setting:

• Theme and default provider/agent model selection
• Editable API provider and per-model override lists
• Editable agent definitions, prompts, tool access, model/thinking overrides, and compaction
• OpenAI Codex login and connection status

Instructions in AGENTS.md

Standing guidance for every session goes in ~/.amber/AGENTS.md. Amber appends the file to the system prompt in a delimited block, so the model can tell your rules apart from Amber's built-in prompt, and treats them as taking precedence wherever the two conflict. Use it for coding conventions, writing style, tools to avoid, and anything else you would otherwise repeat in every session.

# House rules

- Never use the em dash. Use a plain dash instead.
- Run the full test suite before claiming work is done.

A repository can also ship its own AGENTS.md; Amber reads the nearest one from the session's working directory upward and appends it after the user block in a block (skipping the global file).

Both files are optional and are read once when a session starts, so the prompt stays stable for the whole session and survives compaction. An empty or unreadable AGENTS.md is reported on the console once and the session continues without it.

Plugins

Skills and commands can be installed from named marketplaces with the /plugin command, using the same manifest format Claude Code uses, so a marketplace written for Claude Code works unchanged.

/plugin marketplace add anthropics/claude-plugins-official
/plugin list
/plugin install superpowers          # prints the source, ref, version, and commit; installs nothing
/plugin install superpowers --yes    # fetches and installs

An installed plugin is enabled by default and contributes its skills namespaced under the plugin name (superpowers:brainstorming). Bundles live in ~/.amber/plugins/cache, enable state in the [enabled_plugins] table of ~/.amber/settings.toml (or a project's .amber/settings.toml), and /plugin update reports and applies drift against what a marketplace publishes today.

A plugin's skills run shell commands with your privileges. Install only from a source you trust. Install and update both show the resolved source and commit and do nothing until you confirm with --yes.

See docs/plugins.md for the full command reference, the trust model, scopes, and how to author and publish a plugin.

Dependencies

AMBER intentionally has only four direct runtime npm dependencies:

• diff — unified diff generation
• markdown-it — safe Markdown rendering
• smol-toml — settings.toml parsing
• yaml — skill SKILL.md frontmatter parsing

Build and development dependencies are also small:

• typescript
• @types/node
• @yao-pkg/pkg — standalone binary packaging

The packaged binary includes everything needed to run AMBER; Node.js and npm are not required on the destination machine.

Optional: ripgrep

The Grep and Glob tools search with ripgrep when it is installed and fall back to the system grep otherwise, so it is not a hard requirement. Installing ripgrep and keeping its rg binary on PATH is recommended for the best search performance; multiline pattern matching requires it.

brew install ripgrep        # macOS
sudo apt install ripgrep    # Debian/Ubuntu
sudo dnf install ripgrep    # Fedora
sudo pacman -S ripgrep      # Arch Linux
cargo install ripgrep       # any platform with Rust

Useful commands

npm test                 # Build and run all tests
npm run typecheck        # Type-check without emitting files
npm run package:macos    # Build the Apple Silicon binary
npm run package:linux    # Build the Linux x86_64 binary

AMBER binds to 127.0.0.1:3000 by default and is intended for local, single-user use.