Agent + skill

The primary way to use ABX is to let a coding agent run the toolkit. You describe the project in plain language; the agent plans it, deploys it, and stands up its services on infrastructure you own.

Works with any agent

ABX is agent-neutral by design. The CLI is a plain binary any agent with a shell can drive, and the skill is built on the open, cross-agent Agent Skills standard — the same SKILL.md format read by Claude Code, Cursor, Codex CLI, GitHub Copilot, and Gemini CLI. Nothing about the toolkit is coupled to one vendor: the same commands, the same skill, the same update check apply everywhere.

The skill gives an agent the context to select commands, custody bytes, deploy the contract, configure a sale, and run the resolver and effects runner. It asks you the decisions that are yours: which host, how permanent the storage should be, and what a mint should cost.

Install the skill

The skill ships bundled inside the CLI, and installing from the CLI is the recommended path — the bundled copy is version-locked to the CLI it ships with (see Version-locked below), so the guidance the agent reads always matches the commands the CLI actually has.

abx skill install

By default this installs to the two directories that, between them, every supported agent reads:

So one command covers the whole ecosystem. To narrow it to a single agent, pass --agent:

abx skill install --agent claude     # or: cursor · codex · gemini · copilot

Other flags: --global installs under your home directory (~) instead of the current project; --target <dir> writes the skill folder straight under <dir> (an escape hatch for a location we don't special-case). abx skill path prints the path to the bundled skill. After installing, restart your agent so it loads the skill.

Once the ABX repository is public, agents that ship their own Agent Skills installer can also pull the skill straight from GitHub — for example npx skills add ArtBlocks/abx --skill abx-self-host, or an agent-native command like codex skills add …. (That skills installer, incidentally, uses the same .agents/skills hub this command does.) This is a convenience, not the primary path: a git-sourced copy is not version-locked to your local CLI, so prefer abx skill install and use abx doctor to confirm they match.

Version-locked to the CLI

The skill and the CLI are co-versioned — the skill's frontmatter carries the exact version it was cut with (metadata.version in SKILL.md), stamped at release to equal the CLI's version. Because the version travels inside the skill file, abx can tell when an installed copy has fallen behind — regardless of how it was installed — and nudges you to refresh it:

⚠  installed abx skill is v0.1.0-alpha.0 but the CLI is v0.1.0-alpha.1 — refresh it: abx skill install

abx doctor reports the same match explicitly. When you upgrade the CLI, re-run abx skill install to resync — see Upgrading.

Signing

The agent runs the same commands you would, in one of the three signing lanes. On testnet it can use --send with an environment key. For anything with real value, it uses --sign, so a human approves each transaction in a browser wallet, and --dry-run to preview a write first.