CLI, skill, or MCP
Agents reach ZynoHosting through one of two interfaces. Use the CLI with the ZynoHosting skill wherever the agent has a shell. Use the MCP server only where it doesn't: the Claude desktop app.
| Client | Use | Why |
|---|---|---|
| Claude Code (terminal, IDE, desktop Code tab), Codex | CLI + skill | Has a shell. Nothing runs or loads until a task needs it. |
| Muse Code, Cursor, and other local coding agents | CLI + skill | Same as above. |
| Grok Bot, ChatGPT dots | CLI + skill, run on their cloud computer | Sign-in is a link and code the user approves anywhere. |
| Claude desktop app, Chat and Cowork tabs | Desktop extension (.mcpb, the MCP server) | No shell, so a CLI is unusable there. |
| CI and scripts | CLI with ZYNOHOSTING_DEPLOY_TOKEN | No agent involved. A deploy token deploys one site only. |
Why the CLI is the default for coding agents
An MCP server is configured per project and starts with every agent session. It spawns npx at startup and adds over twenty tool schemas and its instructions to the agent's context, even in sessions that never touch hosting. Adding one also requires restarting the agent and approving the server.
A skill costs a one-line description until the agent decides it is relevant, then loads its instructions. The CLI runs only when invoked. The agent can install the skill and use it in the same session, so there is no restart.
The CLI already has what an agent needs:
- the same operations as the MCP tools: sites, files, deploy, preview, promote, download, and ZynoForms (
zynohosting forms …, JSON in and out, with the same inputs and results as the MCP form tools); --jsonwith stable error codes and exit codes;--dry-runfordeploy,promote, andsites set;- a non-interactive guard:
deploy,promote,sites set,download, andfiles rmexit withconfirmation_requiredinstead of prompting when--yes(or--dry-run) is missing.files putapplies immediately, so the skill tells agents to get approval first.
The policy the MCP server delivers as server instructions, such as never guessing a domain or never forcing a promotion, lives in the skill instead.
Why the MCP server stays
The Claude desktop app's Chat and Cowork tabs cannot run a CLI. They are the main non-technical audience, and the signed desktop extension is their only way in. It needs no Node.js install and updates itself.
For Codex and Claude Code, the project-level MCP setup (mcp install) is now optional. Use it only when a team prefers MCP tools to shell commands.
What the CLI does not do that the MCP server does
| MCP behavior | CLI equivalent |
|---|---|
plan_deploy returns a planId; apply_deploy applies exactly that plan. | deploy --dry-run shows a plan; deploy --yes plans again and applies at once. The agent re-runs the dry run if anything changed. |
login returns the one-time URL, or a link and code, as a tool result. | login opens the browser and blocks for up to 300 seconds. Where no browser can open it shows a link and code. For shells that show output only when a command exits, the agent runs login --device --no-wait --json, then login --resume --json. |
list_logins, use_login, and a tenant input on every tool. | logins, logins use, and --tenant (or ZYNOHOSTING_TENANT). |
Local roots (--root, ZYNOHOSTING_MCP_ROOTS) confine local paths. | None. The agent's own permission system governs the shell. |
Possible follow-ups
These would close the remaining gaps between the CLI and the MCP server:
- A plan file for
deployandpromote(for example--plan-out/--plan-in), so an approved dry run is exactly what gets applied. - A
zynohosting skill install --agent claude|codexcommand, so the skill ships in the npm package and stays version-matched to the CLI.
Where the guidance lives
| Interface | Agent guidance |
|---|---|
| CLI | skills/zynohosting/SKILL.md, installed by install.md |
| MCP | SERVER_INSTRUCTIONS and tool descriptions in src/mcp/server.ts and src/mcp/forms.ts, delivered on every connection |
Keep both in step when behavior changes. The preview guide is the shared reference for the preview model.