mobius-1

Interactive CLI for chatting with Claude via the Agent SDK (V2 interface).

Mobius is an autonomous AI engineer that runs in your terminal. It uses Anthropic's Agent SDK to give Claude full agency: it can write code, run commands, browse the web, and deploy—all without you in the loop. You give it a goal, and it works toward it turn after turn, making real progress on real products.

What makes Mobius different is that it's built for marathon runs. Unlike one-shot chat interfaces, Mobius persists its state to disk. Stop it, restart it, close your laptop—when you come back, it resumes exactly where it left off with full context. The agent runs in a self-reprompting loop that keeps going until the job is done. You can let it run overnight, across days, or in a Modal sandbox in the cloud. It's an AI that actually finishes things.

The project supports both local execution and a sandboxed "safe" mode on Modal, so you can run agent workflows in an isolated environment and stream results back to your terminal. A generic Modal wrapper also lets you run arbitrary commands in the cloud with optional port tunneling.

Setup

bun install

Make sure your .env has the required credentials configured.

Run On Desktop (default)

bun run agent

Run Safe Copy On Modal CPU

bun run agent:safe

Install/auth once for safe mode:

python3 -m pip install modal
modal token set --token-id "$MODAL_TOKEN_ID" --token-secret "$MODAL_TOKEN_SECRET"

Safe mode runs the same command path in a Modal sandbox and streams terminal output live.

Generic Modal Wrapper

Run any command in Modal:

./modal/run -- <your command>

Examples:

./modal/run -- bun run index.ts
./modal/run --port 3000 -- npm run dev

If a port is exposed, the script prints the Modal tunnel URL.

Configuration

Variable	Default	Description
CLAUDE_MODEL	claude-sonnet-4-5-20250514	Model to use