agent-kit

@falkenslab/agent-kit is scaffolding for building agents on top of the Claude Agent SDK (@anthropic-ai/claude-agent-sdk). It isn't an agent: it solves once what every agent needs, so a concrete agent only writes what belongs to its own domain.

What the kit gives an agent

Area	What you get
Oversight	Three modes: autonomous, guided (an approval before anything visible or hard to undo) and interactive (an approval before every tool call). Guided and interactive switch into each other live.
Human in the loop	Approval and manual-intervention checkpoints, answered from the keyboard, from a panel in the Ink UI, from a desktop app through an InteractionPort, or by writing a response file.
Safety	Hooks that keep the file tools inside the agent's own folders, protect paths, keep Bash for subagents only, allow only the subagents you declare and run them in the foreground.
Memory	A built-in knowledge base (an "LLM wiki" the agent maintains), originals kept untouched in a sources folder, and every conversation stored in its run folder to resume it later.
Tools	Your own MCP servers (in-process or external), skills and slash commands in plugins, and subagents.
Terminal UI	A full screen Ink chat modelled on Claude Code, a plain readline chat, a progress view for one-shot runs and a question wizard, with themes and four languages.
Hosts	A core that never touches the terminal, a normalized event stream (runQuery()), and ports so a desktop app or a server can drive the same agent.
Trail	A session log of what the terminal showed and a transcript of every tool call with secrets redacted.

What an agent writes itself

An agent implements one small contract, AgentSpec: its system prompt, the MCP servers it always has, the folders with its skills and commands, and its subagents. Everything else is configuration:

• a config object (BaseSessionConfig, extended with the agent's own fields) with the mode and the folders it works in;
• the options of the UI it wants (a header, a welcome message, a theme, its own texts).

buildSessionOptions() turns the spec and the config into the SDK's Options for query(), and a chat or a progress view runs the session.

How this documentation is organized

1. Getting started: install the kit, authenticate, and build a first agent step by step.
2. Core concepts: the architecture, AgentSpec, the session config, the modes and exactly what buildSessionOptions() builds.
3. Capabilities: tools and MCP servers, skills and plugins, subagents, the knowledge base and prompt files.
4. Human in the loop: checkpoints, the response file and custom interaction ports.
5. Security: every guardrail and why it exists.
6. Terminal UI: the chats, the progress view, the wizard, tool labels and themes.
7. Sessions: run folders and resuming, languages, context size and cost.
8. Advanced: the event stream, one-shot runs, hosts without a terminal and your own hooks.
9. Examples: Captain Whiskers, file by file, and short recipes.
10. Reference: SDK behaviors the kit relies on, troubleshooting and versioning.
11. API reference: every export, generated from the source.

Requirements

Node.js 20 or later, and a Claude subscription token or an Anthropic API key. The kit is ESM only and written in TypeScript; its type declarations ship with the package.