dokploy-ghcr-deploy（dokploy ghcr 部署）

Dokploy + GHCR auto-deploy

End-to-end recipe for turning a Dockerized GitHub repo into a Dokploy app that redeploys automatically on every push to main. Builds happen in GitHub Actions; images go to GHCR (private by default); Dokploy pulls and redeploys via a REST call from the workflow's final step.

Why this flow:

• GHCR over Docker Hub — Docker Hub's free tier caps at 1 private repo, GHCR gives unlimited private packages. Push-side auth is the built-in GITHUB_TOKEN, no extra PAT needed on the workflow side.
• API-trigger over registry webhook — GHCR doesn't fire push webhooks, so the workflow POSTs to Dokploy's /api/application.deploy instead of relying on a registry-side webhook (which is how Docker Hub → Dokploy would work).

Work through the phases in order. Each one has a clear stop condition; don't skip ahead if something's red.

Prerequisites — check before anything else

Do these silently in parallel where possible; only surface problems to the user.

1. Dockerfile exists and exposes a port. Read it. Grep for EXPOSE. Capture the port number — you'll need it for the Dokploy domain. If there's no Dockerfile or no EXPOSE, stop and ask the user; do not fabricate one from framework defaults (that would be silent guesswork).
2. .dockerignore exists. If missing, offer to add a minimal one (node_modules, .git, build, dist, tests, .env*). Not a blocker.
3. gh CLI authenticated. gh auth status should show an active account. If not, stop and ask the user to run gh auth login first.
4. Dokploy MCP is connected. Call mcp__dokploy-mcp__user-session. If it errors, stop and ask the user to connect the Dokploy MCP before continuing — the skill can't do the Dokploy half without it.
5. GitHub remote. Parse git remote get-url origin to extract owner/repo. Lowercase it — GHCR rejects uppercase in image paths. If the remote isn't GitHub, this skill doesn't apply; tell the user.

Inputs to gather from the user

Batch these into one AskUserQuestion call so the user answers once, not five times. Defaults in parentheses — pre-fill them and let the user override:

• Dokploy project name (default: the repo name)
• Dokploy app name (default: web)
• Public domain to attach (must already resolve to the Dokploy host; if they have a wildcard like *.example.com, suggest <project>.example.com)
• Dokploy base URL (e.g. https://dokploy.example.com) — you cannot infer this; always ask
• Extra env vars beyond APP_ENV=production — offer to skip; they can add more in the Dokploy UI later. Validate each entry matches ^[A-Z_][A-Z0-9_]*= and reject values containing raw newlines before sending to application-saveEnvironment, so a stray newline in one value can't silently inject another variable.

Phase A — Repo prep

1. Write .github/workflows/docker-publish.yml from assets/docker-publish.yml.template. The template uses ${{github.repository}} and lowercases it at runtime, so it works unmodified for most repos — you don't normally need to substitute anything.
2. Create a feature branch if not already on one (don't commit straight to main — we want the workflow to land atomically with a PR once secrets are in place).
3. Stage and commit the workflow. Don't push yet — we want Dokploy and GitHub secrets configured first, otherwise the very first workflow run will fire against an app that isn't ready and leave a failed deployment in the history.

Phase B — Dokploy setup via MCP

Sequential, fail-fast. Each tool call produces IDs you'll need for the next one — capture them as you go.

1. Resolve the organization — call mcp__dokploy-mcp__organization-active and save organizationId.
2. mcp__dokploy-mcp__project-create with name and a short description. The response includes both a projectId and an auto-created default environmentId (named production) — save both.
3. mcp__dokploy-mcp__application-create with name (the app name) and environmentId. Save applicationId. The app starts with sourceType: "github"; it flips to docker when you save the Docker provider next.
4. mcp__dokploy-mcp__application-saveDockerProvider with:

- applicationId - dockerImage: "ghcr.io/<lowercase-owner>/<lowercase-repo>:latest" - registryUrl: "ghcr.io" - username: "<lowercase-owner>" - password: "PAT_PLACEHOLDER" — this stub keeps the MCP call valid and is left in place. The real PAT is set by the user directly in the Dokploy UI in Phase E, so it never transits this conversation.

1. mcp__dokploy-mcp__application-saveEnvironment with:

- applicationId - createEnvFile: true - env: "APP_ENV=production\n..." (plus whatever the user added) - buildArgs: "", buildSecrets: "" — explicit empty strings, not undefined. See Gotchas.

1. mcp__dokploy-mcp__domain-create with:

- applicationId - host: "<domain-the-user-gave>" - port: <port-from-Dockerfile> - https: true, certificateType: "letsencrypt" - path: "/", stripPath: false, domainType: "application"

1. mcp__dokploy-mcp__user-createApiKey with name: "github-actions-deploy-<repo>" and metadata: {"organizationId": "<from step 1>"}. Save the key — it's shown once. Flag to the user that the default rate limit is 10 requests / 24h; offer to raise it via the Dokploy UI if they push to main a lot.

Phase C — GitHub secrets via gh

Three secrets, set in one batched Bash call so the user approves once:

gh secret set DOKPLOY_URL --repo <owner>/<repo> --body "<dokploy-base-url>" && \
gh secret set DOKPLOY_API_KEY --repo <owner>/<repo> --body "<api-key-from-B7>" && \
gh secret set DOKPLOY_APPLICATION_ID --repo <owner>/<repo> --body "<applicationId-from-B3>"

Then smoke-test the API before pushing code. read -s keeps the key out of ~/.bash_history; the unset at the end clears it from the shell so it doesn't linger in the process environment:

read -rsp "Paste DOKPLOY_API_KEY (not echoed): " DOKPLOY_API_KEY; echo
DOKPLOY_URL='https://your-dokploy.example.com'
APP_ID='<applicationId from B3>'
DOKPLOY_API_KEY="$DOKPLOY_API_KEY" \
  curl -fsS --max-time 30 -o /dev/null -w "%{http_code}\n" \
  -X POST "${DOKPLOY_URL%/}/api/application.deploy" \
  -H "x-api-key: $DOKPLOY_API_KEY" \
  -H "Content-Type: application/json" \
  --data "$(jq -cn --arg id "$APP_ID" '{applicationId:$id}')"
unset DOKPLOY_API_KEY

Expect 200. Anything else means the URL, key, or app ID is wrong — fix before moving on. The deploy will fail (no image has been pushed yet) but the trigger should succeed; this is proof the auth path is good.

Phase D — Push & PR

Push the branch, open a PR with gh pr create. PR body should call out:

• the three secrets that must be set (confirm they're in place)
• the Dokploy app details (project, app, domain, port)
• that merging triggers the first real deploy

Phase E — GHCR PAT (the only manual step)

GitHub has no API for minting PATs, so the user has to do this in the UI. Don't pretend otherwise.

The PAT must not transit this conversation. A read:packages PAT in chat history is credential exposure — transcripts are routinely retained server-side, and the PAT can pull every private package the user can see. So the user pastes it straight into Dokploy, not into chat. Do not ask them to paste it here, and do not call mcp__dokploy-mcp__application-saveDockerProvider with the real value; the PAT_PLACEHOLDER stub from Phase B step 4 is replaced in-place via the Dokploy UI.

Tell the user verbatim:

1. GitHub → Settings → Developer settings → Personal access tokens (classic) → Generate new token → scope read:packages only → set expiration to 90 days (GitHub's recommended max for machine credentials; do not pick "No expiration") → generate → copy the token to your clipboard. 2. Dokploy → Projects → *your project* → *your app* → General → Docker provider → paste the token into the Password field → Save. 3. Reply here with "saved" (no token, no paste).

Only when they confirm "saved" do you move to Phase F.

Rotation, if it's ever needed, follows the same path: revoke in GitHub, generate a new PAT, paste it into the Dokploy UI. No chat step.

Why this step isn't automated: the gh CLI token doesn't carry read:packages by default (scopes are typically repo, read:org, gist, admin:public_key), and running gh auth refresh -s read:packages would give you a user-scoped OAuth token with every other scope still attached — too much privilege to stash as a service password. A dedicated classic PAT scoped to read:packages, set via the Dokploy UI, is the cleanest answer.

Phase F — First deploy

1. Merge the PR (or push to main). The workflow fires.
2. Watch the Actions tab: the build should push two tags (latest and sha-<short>) and the final step should return 2xx from Dokploy.
3. In Dokploy → app → Deployments, confirm a new deployment went from "Pulling" → "Running" → "Done".
4. Hit the domain; expect the app to respond on the port from the Dockerfile.

If the Dokploy deployment shows "denied: denied" from ghcr.io/v2/, the PAT is wrong or expired — go back to Phase E.

Gotchas (learned the hard way)

• MCP zod unions reject undefined. application-saveDockerProvider requires all of registryUrl, username, password present (pass empty strings or real values, not omitted). application-saveEnvironment requires buildArgs and buildSecrets — pass "". Skipping them wastes a round trip on an opaque validation error.
• GHCR image paths must be all-lowercase. The workflow template handles this via ${IMAGE_NAME,,}. If you hand-roll one, don't forget.
• Don't suggest the Docker Hub webhook flow for GHCR. GHCR doesn't support push webhooks. The Dokploy docs mention the webhook flow in the Docker Hub context only.
• Docker Hub free tier = 1 private repo. That's why this skill defaults to GHCR. Only mention Docker Hub to explain why you're not using it.
• Package visibility. GHCR packages inherit visibility from the source repo on first push — private repo → private package. No extra API call needed. If the repo is public, the package will be public too; mention that to the user once.
• Dokploy API key default rate limit is 10/24h. Fine for most push to main flows. Raise in the Dokploy UI if the user does many deploys per day (e.g., trunk-based with many merges).
• gh auth status scopes aren't enough for GHCR pulls. Typical gh tokens carry repo, read:org, gist, admin:public_key — no read:packages. That's why Phase E can't lean on the CLI token.
• Don't rename APP_ENV. Codebases that ship dual Vercel + Dokploy deploys often branch on APP_ENV specifically to avoid tangling with VERCEL_ENV. Set it to production unconditionally on Dokploy.

What this skill deliberately doesn't do

• Doesn't migrate an existing Dokploy app's domains — if the user is replacing a GitHub-source Dokploy app, ask whether to create a new app in a new project (clean, no downtime on the old one) or reconfigure in place (needs domain cutover). Default to new project.
• Doesn't copy env vars from one app to another. Listing them is fine; blindly copying live Stripe/DB secrets between apps risks clobbering or double-writing.
• Doesn't set up preview deployments, healthchecks, or replicas. Scope creep; add when the user asks.