Files
Containr/docs/guides/CLI.md
T

177 lines
6.6 KiB
Markdown

# Containr CLI
The CLI controls every part of a Containr instance — the same API surface
the web UI uses. Build it with `scripts/build-cli.sh` or
`go build ./cmd/cli` from `app/backend`.
## Authentication
Create a personal access token in the web UI under
**Settings → Personal Access Tokens** (or `containr tokens create`), then:
```bash
containr auth login cnp_... # verify + store
containr auth login cnp_... --url https://containr.example.com
containr auth login cnp_... --name production # named profile
containr auth status # show active profile + identity
containr auth logout
```
Resolution order for every command: `--token`/`--api-url` flags >
`CONTAINR_TOKEN`/`CONTAINR_API_URL` env > active profile >
`http://localhost:8080/api/v1`. The active profile is `--profile` >
`CONTAINR_PROFILE` > `current_profile` in `~/.containr.yaml`.
## Output and scripting
- `--json` prints the raw API response — safe to pipe and parse.
- `--yes` / `-y` skips confirmation on destructive commands (delete,
revoke, restore, failover). Non-interactive shells refuse destructive
commands without it.
- Exit codes are stable: `0` ok, `1` error, `2` auth/permission,
`3` not found, `4` validation, `5` unreachable.
## Commands
```
auth login|logout|status token auth and profiles
tokens list|create|revoke personal access tokens
projects list|get|create|delete
services list|get|create|update|delete|start|stop|restart|redeploy
services domains list|add|remove|default|check multi-domain + DNS preflight
volumes list|delete docker volume inventory + cleanup (admin)
deploy <service-id> trigger a deployment (--no-cache, --commit-hash, --branch)
deployments list|get|logs|rollback|cancel
logs <service-id> [--follow] runtime logs (SSE follow)
exec <service-id> -- <cmd> one-off container command
variables list|set|unset env vars (secrets masked; --redeploy applies them)
variables shared list|set|unset project shared vars — referenced as ${{shared.KEY}}
databases list|get|create|delete|action|backup|restore|download-backup
cron list|get|create|delete|trigger|executions
templates list|get|create|delete|deploy
nodes list|get|delete|tokens node agents + onboarding tokens (admin)
scaling status|policies|services|scale
ha status|alerts|health|policies|enable|disable|failover
security scan|vulnerabilities|metrics
gateway services|keys|analytics (admin)
admin overview|users|settings|audit-logs
notifications list|read
up deploy the current directory
```
## containr up
Run inside a git checkout: uses `remote.origin.url` + the current branch
as the build source, creates a service named after the directory when none
exists, then triggers a deployment.
```bash
containr up --project <id>
containr up --project <id> --image ghcr.io/example/app:latest # image, no build
```
Set `project_id` on the active profile to omit `--project` (stored in
`~/.containr.yaml` under `profiles.<name>.project_id`).
## Service volumes
`services update` manages mounts; flags replace the whole mount list:
```bash
containr services update <id> --volume data:/data --bind /srv/cfg:/etc/app:ro
containr services update <id> --clear-volumes
```
`type` is `volume` (docker named volume, auto-created) or `bind` (host
path). `target` must be absolute. Mounts apply on the next deploy or
redeploy — `containr deploy <id>` to apply immediately.
`volumes list` shows the node's docker volume inventory with in-use
flags; `volumes delete <name>` removes an unused volume (admin scope).
## Service domains and access
A service can carry multiple hostnames; the default is mirrored to
`services.domain`:
```bash
containr services domains add <id> app.example.com --default
containr services domains check <id> # DNS preflight per domain
containr services domains remove <id> <domain-id>
```
DNS check compares each domain to the expected target — set
`app_settings.public_ip` (or `PUBLIC_IP` env) so `wrong-target` is
meaningful; without it any resolvable domain reports `ok`.
Access gates apply via Traefik middlewares (no extra proxy):
```bash
containr services update <id> --maintenance on
containr services update <id> --basic-auth admin:secret --basic-auth ops:pw2
```
Maintenance mode redirects all traffic to `<BASE_URL>/api/v1/maintenance`
— set `app_settings.base_url` or `BASE_URL`; without it the service is
taken offline with empty responses. Basic-auth credentials are stored
bcrypt-hashed (htpasswd) and gate every request.
## Build strategy and resources
Source-built services pick a builder per service:
```bash
containr services update <id> --builder railpack # auto|railpack|nixpacks|dockerfile|static
containr services update <id> --builder static --static-cmd "npm run build" --static-dir dist
```
`auto` detects from repo contents (Dockerfile → railpack → nixpacks).
`static` builds assets in a node stage and serves them from nginx —
`--static-cmd` runs inside the build stage (default
`npm ci && npm run build`), `--static-dir` is copied into the nginx image
(default `dist`, exposed on port 80).
Resource reservations complement the hard `cpu`/`memory` limits:
```bash
containr services update <id> --cpu-reserve 0.25 --memory-reserve 128Mi
```
`cpu`/`memory` are hard caps (NanoCPUs/Memory); `cpu_reserve`/
`memory_reserve` are soft hints (CPU shares, MemoryReservation). Empty
string clears. `app_settings.default_cpu`/`default_memory` change the
instance-level defaults used at service creation;
`app_settings.capacity_policy=block` fails deployments whose requested
memory exceeds node capacity (default `warn` logs only).
## MCP server
`containr-mcp` (`go build ./cmd/mcp`) exposes the same API as MCP tools
over stdio — every command group above maps to a `containr_*` tool
(~60 tools). Destructive tools require the caller to pass
`confirm: true`; results are pretty JSON with `isError` on failures.
Client configuration (Claude Code, Cursor, Devin, any stdio MCP host):
```json
{
"mcpServers": {
"containr": {
"command": "containr-mcp",
"env": {
"CONTAINR_API_URL": "https://containr.example.com",
"CONTAINR_TOKEN": "cnp_..."
}
}
}
}
```
Scope the token deliberately: `read` is enough for inspection agents,
`write` for deploy automation, `admin` only for platform operations.
Deployments are serialized per service: a second `deploy` on a busy
service lands in `queued` state and runs when the active one finishes.
`deployments cancel <id>` drops a queued deployment or aborts a running
one (builds/reconciles stop via context cancellation).