@chrischall/mcp-host-cli

mcp-host — the admin CLI for an mcp-host deployment: register npm / GitHub / PyPI / MCPB servers, manage secrets, and inspect, restart or remove hosted MCPs through the gateway control API.

npm i -g @chrischall/mcp-host-cli      # Node 20 or newer; no runtime dependencies

export MCP_HOST_URL=https://mcp.nullnet.app
export MCP_HOST_ADMIN_TOKEN=...

mcp-host register                       # no flags on a terminal: the step-by-step wizard
mcp-host list
mcp-host help                           # every command

One-off use without installing: npx @chrischall/mcp-host-cli <command>.

The CLI only talks to the gateway named by MCP_HOST_URL, and refuses an http:// URL unless --insecure is passed, because the request carries your bearer token.

Its version tracks the mcp-host release it was cut from, and its commands track that release's control API — keep it current with the deployment you point it at.

Upgrading from 2.33.0

The release after 2.33.0 is stricter in two ways a script can notice. Both turn a silent mistake into an error, so they ship as a fix rather than a major version:

• Every command refuses a flag it does not take. Before, ten commands (set, register, connector, build, preview, bundles, infra, secret, marketplace, bridge) dropped an unknown flag and ran anyway — set reg_1 --denny destructive sent the update without the policy you typed. Now it exits 1 with unknown flag --denny and the command's usage. A script that passed a flag the command never read (for example preview --follow, secret rm --legacy, infra sync --days) must drop it.
• approvals approve <id> asks before it approves. It shows the tool, the registration and the arguments, then needs a "y" at a terminal. Without a terminal (a pipe, CI, a cron job) it approves nothing and exits 1 unless you pass --yes, so add --yes to any script that approves calls.

An unreachable gateway now prints one mcp-host: line and exits 1 instead of a fetch failed stack trace.