Features · TypeScript Coding Agent

A capability summary of the indusagi-coding-agent CLI, derived from the subsystems that ship in the source. Everything here is built from scratch on the single indusagi framework dependency.

At a glance

Three run modes: interactive REPL, one-shot print, and a JSON line protocol.
A built-in tool deck: file read/write/edit, search, shell, background processes, web, and a task checklist.
External tools over MCP — connect to MCP servers and graft their tools into the deck.
Multi-provider models plus optional external runtimes (CLI-backed agents).
Loadable skills and subagents, slash commands, transcript export, and tracing.

Run modes

Invocation	Mode	What it does
`indus`	interactive	Ink/React REPL — streaming replies, a live tool deck, slash commands
`indus -p "…"` / `--print`	print	One request, prints the final result, exits
`indus --json` / `--rpc`	line protocol	Newline-delimited JSON for a parent process to drive
`indus signin` / `signout`	credentials	Store / clear an API key in the local auth vault

indus --help / -h prints the generated flag reference; indus --version / -v prints the version. --interactive / -i forces the REPL even when a prompt is supplied. See JSON Mode and RPC Mode.

Built-in tools

The capability deck exposes the framework's built-ins as agent tools, in catalog order. Read-only tools are available in every profile; mutating tools require an authoring (full-access) session.

Tool	What it does	Access
`read`	Return a file's contents (optional offset + line cap; renders images inline where supported)	read-only
`ls`	List a directory's entries (optional cap)	read-only
`grep`	Scan files for lines matching a pattern (case, literal, context, cap)	read-only
`find`	Locate files/directories by glob-style name beneath a root	read-only
`websearch`	Query the live web and return ranked result snippets	read-only
`webfetch`	Retrieve a URL and return its body as text, Markdown, or HTML	read-only
`todoread`	Read back the session's task checklist	read-only
`write`	Create or overwrite a file (makes parent dirs)	mutating
`edit`	Find-and-replace an exact span in a file and report the diff	mutating
`bash`	Run a shell command in the working directory, streaming output, with an optional timeout	mutating
`process`	Start, list, inspect, feed, and stop long-running background commands	mutating
`todowrite`	Replace or amend the session's task checklist	mutating

Limit a run to a subset with --tools read,grep,bash (comma-separated or repeated), or disable every built-in with --no-tools. The bash tool runs in a non-interactive shell — see Shell Aliases.

MCP — external tools

The agent is a client of external Model Context Protocol servers: it connects to each configured server, lists its tools, and grafts them into the deck under qualified "<server>__<tool>" names. Enrollment is event-sourced, so servers can attach and detach over a session's life without mutating shared state. Attach servers from the command line with --mcp <endpoint>, manage them in an interactive session with /mcp, or declare them in a config file. See MCP.

Models and runtimes

Multi-provider. The session defaults to anthropic (defaultProvider), with the model picked by defaultModel and reasoning effort by defaultThinkingLevel. Choose a model per run with --model / -m (provider-qualified or bare) and reasoning with --thinking. List models with --list-models. See Providers and Models.
External runtimes (provider routing). Beyond the framework's network stream, the runtime-bridge can route a turn to a child coding-agent process driven over its own protocol. Built-in adapters cover an Anthropic-flavoured CLI (claude-cli), an OpenAI-flavoured CLI (codex-cli), and a peer indusagi agent (indusagi-cli). A model is annotated with an external-runtime spec; a model with no spec routes normally. A CLI runtime that owns its own auth can be offered even with no API key on disk.

Authentication

Set ANTHROPIC_API_KEY (or another supported provider key) in the environment, or run indus signin. Two methods are supported: an API key or a browser sign-in (OAuth) for the providers that offer it (Anthropic, OpenAI Codex, GitHub Copilot). indus signin prefers the browser path for an OAuth-capable provider and falls back to an API key; force one with --method oauth / --method api-key (shorthands --oauth / --api-key). In an interactive session use /login. Authenticate a specific run with a stored account via --account. Credentials live in the local auth vault under the profile directory.

Working memory

A single working-memory tool lets the agent carry a persistent scratch note across turns within a session: read, replace, and append actions over one text buffer. It is a plain in-process note — not a vector or semantic store. Inspect or toggle it with /memory.

Composio (SaaS bridges)

When COMPOSIO_API_KEY is set, /composio connects, enables, and inspects Composio toolkits, exposing their remote tools to the agent through the SaaS gateway. Without the key, the command reports the missing-key state explicitly.

Extensibility

Skills — on-demand capability packages discovered and (optionally) surfaced as their own slash commands. See Skills.
Subagents — spawn specialized agents for delegated work. See Subagents.
Addons / extensions — loadable modules that contribute tools and commands and hook lifecycle events; install sources via the extensionPackages setting. See Extensions and Loading Extensions.
Themes & keybindings — name a colour scheme with colourScheme and rebind console keys. See Themes and Keybindings.

Sessions and export

Sessions are persisted under the profile directory; resume one with --resume / -r or continue the latest in the current directory with --continue / -c. In a session, /copy puts the last reply on the clipboard, /export [path] renders the transcript to standalone HTML, and /share publishes it as a secret GitHub gist via the gh CLI. See Session Format.

Observability

The insight subsystem provides tracing with sinks, replay, and secret redaction. Set INDUSAGI_DEBUG in the environment to surface debug output; /debug toggles the reasoning display and dumps a JSONL snapshot of the session (stats plus every message) to a temporary file, reporting the path it wrote.