Languages and tools

What every sandbox has before its first command: the languages, build tools, coding agents and assistants a task asks for most.

Every sandbox starts from the same image: Ubuntu 24.04 on x86-64, with what follows installed and on the PATH a command is given. Nothing has to be installed first, and anything else can be: commands run as root.

Languages#

Tools#

Coding agents#

Nine coding agents are installed. Each signs in to its model with your own key. Keep the key as a secret of the sandbox's template, under the name the tool reads, such as ANTHROPIC_API_KEY for Claude Code, and it reaches every command.

cd /opt/agental/app && claude -p "make the failing test pass" --dangerously-skip-permissions

• Commands run as root, and Claude Code skips its permission prompts as root only when it is told it is in a sandbox. claude here tells it so: the machine is yours alone.
• An agent run with -p or exec answers and ends. For a longer task, start it in the background as in Running commands.
• The agents with a version in the table are pinned at it, and each file they came from was checked against a sum when the image was built. Cursor's CLI does not update itself here, and Kimi Code does not install its updates.
• Kimi Code reads no key from the environment by itself. Sign in with /login in kimi, or add a provider to /root/.kimi-code/config.toml whose api_key_env names the secret, such as KIMI_API_KEY.

Assistants#

OpenClaw and Hermes are assistants: you talk to one from a chat app such as Telegram, Slack or WhatsApp, and it keeps working between your messages. Each is installed, with its settings and its folder, and its gateway, the program that connects it to chat apps and keeps it running, is off. Nothing starts it until you do.

openclaw onboard
nohup openclaw gateway > /root/openclaw-gateway.log 2>&1 < /dev/null &

• Set one up with openclaw onboard or hermes setup, over SSH or on the desktop: each asks for a model key and the chat apps to connect. Its key can also be a secret of the template, as for the coding agents.
• Start its gateway in the background, as above, or with hermes gateway run. A gateway stops when the sandbox stops, and is started again by you after it starts. To keep an assistant answering, keep its sandbox running.
• Its gateway listens only on the sandbox's own 127.0.0.1 unless you change that, OpenClaw's at port 18789. Most chat apps are reached by a connection the gateway opens itself. One that has to call the gateway needs a public address for its port.
• To have a gateway start by itself, openclaw gateway install and hermes gateway install make it a service, as each one's own docs describe.
• Hermes runs on a Python, Node.js, ripgrep and ffmpeg of its own, which its installer brought. Its browser and its computer-use driver were left out: install them with hermes pm install agent-browser and hermes pm install cua-driver if a task asks for them.

What the agents are told#

Each agent starts already knowing where it is, so a prompt can start with the task. It reads a short note with its standing instructions, every time: that it is root on an agental sandbox, to work in /opt/agental, that the disk outlasts a stop and running processes do not, to put long work in the background, to have a server listen on 0.0.0.0, and that AGENTAL_SANDBOX_ID and AGENTAL_API_URL name its sandbox. Two skills, which it loads when a task asks for them, say more: agental, about the machine and about calling the API on its own sandbox, and playwright-cli, about driving the browser.

• The agental skill calls the API only when you have given the agent a key, as a secret of the template such as AGENTAL_API_KEY, and then only on its own sandbox and the forks it made of it: to give a port an address, to fork the sandbox when you asked it to, run commands in the forks and delete them when it is done, to write to you by email, or to stop the sandbox when you asked it to. Give it a key that can do only that, limited to its sandbox: email.send alone, for an agent that only needs to write to you, or sandbox.fork, exec and sandbox.delete-own for one that forks.
• Your own rules go where the agent finds them on any computer: an AGENTS.md or CLAUDE.md in the project it works in, or more lines in the files above. A snapshot keeps them for the next sandbox.
• A sandbox made from an image built before these were added has neither the note nor the agental skill. The variables are in its commands all the same.

Installing more#

• npm install -g works as it does on any Ubuntu machine, and so does apt-get install -y after apt-get update, which brings the package lists up to date.
• Packages install without their docs, man pages and translations, as in Ubuntu's minimal images. To have them, remove /etc/dpkg/dpkg.cfg.d/01-agental-nodoc and install the package again.
• Ubuntu keeps pip install to virtual environments, and refuses it outside one. Make one with uv venv or python3 -m venv, or install a Python tool with uv tool install.
• A command's PATH puts /root/.local/bin, /root/.cargo/bin, /root/go/bin and /root/.bun/bin first, where uv tool install, cargo install, go install and bun add -g put what they install. The next command finds it, and a newer copy there is used before the image's own.
• To start every new sandbox with what you installed, take a snapshot and have its template start sandboxes from it.
• A sandbox keeps the versions it was made with. A sandbox made after the image is built again has the newer ones.

The full machine-readable reference is OpenAPI 3.1. Every page as Markdown, for an agent: llms.txt. Questions: [email protected]