Install
curl -fsSL https://get.cotal.ai | shWho it's for
Cotal is a provider-agnostic, cross-machine-capable, extensible open standard, positioned to let multiple AI agents work together. If you use more than one of Codex, Cursor, OpenCode, Gemini CLI and Windsurf and want them to share the same set of Agent Skills, or you frequently switch machines and want a workflow that stopped midway to keep running on another host, this is the right tool for you. Claude Code users go through the officially provided cotal-skills plugin instead.
If you only use a single client and don't need to resume tasks across machines, the incremental benefit is limited. The official page does not provide the following information, so treat it as unconfirmed before you decide: supported operating systems and version requirements, whether an account or keys are required, and scenarios where it is explicitly unsuitable. The one boundary that can be confirmed today: installation writes into the user home directory and needs no sudo, which usually implies a Unix-like terminal environment, but the official page does not state this explicitly, and whether Windows works is unconfirmed.
Your first attempt in 3 minutes
There is exactly one recommended main path; the commands and paths below all come from the official documentation.
Prerequisites: a terminal that can run curl; at least one of Codex, Cursor, OpenCode, Gemini CLI or Windsurf installed on this machine (for Claude Code, see the note at the end of this section).
Step one, run the official install command. One confirmation point to spell out first: this command downloads a remote script and hands it to the shell to execute, writing content into your home directory, and no sudo is needed at any point; accept this before you press Enter.
curl -fsSL https://get.cotal.ai | sh
Step two, follow the official guidance and run the post-install setup command. It seeds a default persona and installs Cotal's Agent Skills into ~/.agents/skills.
cotal setup
Step three, open any supported client and launch a low-risk, read-only mini task, for example asking the agent to describe which categories of content the freshly installed Agent Skills contain and what scenarios each fits. The success signals you can observe: the install script ends normally, and the setup command completes the persona seeding and the skills installation. How these skills are presented in each client's interface and how the first conversation is routed is not documented on the official page.
Secondary note: Claude Code uses the official cotal-skills plugin; the official page does not go into installation details.
Use cases
Case 1: multiple clients sharing one skill set. Scenario: you use Cursor and Gemini CLI side by side and don't want to maintain a separate copy of your skills for each tool. Approach: after the installation and setup above, have the agent in one of the clients do something small with the freshly installed skills, for example summarize which Cotal capabilities it now has. Basis: the official project provides Agent Skills for these five clients and installs them into ~/.agents/skills. How each client loads that directory and how content stays in sync is not documented on the official page; the confirmable scope ends here.
Case 2: resuming a stopped workflow on another machine. Scenario: a workflow exits halfway through on your laptop and you want to continue on your desktop. Approach: after completing the same installation on the new host, have the agent resume this run from its journal. Basis: the official documentation states that workflow runs record every step into a journal, and a stopped run can resume from its journal on any host. The exact resume command and the storage location of the journal are not given on the official page; don't guess paths on your own.
Usage boundaries
Files and commands: there are two confirmed write points — the install script writes into the home directory (no sudo), and the setup command writes the default persona and ~/.agents/skills. Beyond that, whether the agent reads or writes other files at runtime, executes commands or accesses the network is not documented on the official page; treat all of it as unknown and don't assume these operations won't happen.
Runtime environment and clients: installation depends on curl in a terminal; for supported clients, defer to the official list — the official page does not state that only these clients are supported, and whether other clients work is unconfirmed.
Official restrictions and unconfirmed items: the official page lists no explicit usage restrictions. Other key facts still unconfirmed include: system requirements, how to uninstall, which skills are actually included, where the journal is stored, and whether data is collected. When you run into these gaps, treat them as unknown and don't fill in conclusions on the project's behalf.