AI agents
Infrastructure an agent can operate without guessing.
Structured JSON on every command, real process exit codes instead of parsed prose, and an explicit gate in front of anything destructive.
A coding agent driving a deployment platform fails in a specific way: it reads human-formatted output, guesses that something worked, and moves on. The wording changes, the guess breaks, and nobody notices until production does.
Dockup's CLI answers in JSON with a documented shape and exits with a real code. An agent does not have to interpret a sentence to know whether a deployment succeeded, and the operations that cannot be undone stop and ask rather than proceeding on a confident-sounding assumption.
- Every command
- Accepts --json and returns a documented structure. The prose output is for people.
- Exit codes
- Real ones. A failed deploy exits non-zero with a code that says where to look.
- Destructive operations
- Behind an explicit confirmation. An agent cannot delete a service by being confident.
- Secrets
- Redacted from output, so a transcript is not a credential leak
- Security
- Image CVE and configuration scan available per service
- Works with
- Claude Code, Codex, Cursor, and anything else that can run a command
Why waiting is a feature
The temptation with agent tooling is to remove every confirmation, because confirmations are friction and friction is what agents are meant to eliminate. The result is a tool that will eventually delete the wrong database very efficiently.
Dockup keeps the gate and makes it machine-readable instead: a refusal returns a code the agent can recognise and surface, rather than a prompt it will try to answer.
Read next
- Deploying to production from Claude CodeClaude Code deployment with Dockup: install the agent skill, authenticate safely, deploy from Git, verify success, and operate production safely.
- End-to-end deployment with CodexCodex deployment with Dockup, from CLI and skill installation to Git service creation, JSON verification, health checks, rollback, and safe retries.
- Designing a CLI for AI agentsAI agent CLI design requires structured JSON, real exit codes, terminal-state waiting, stable errors, and safe confirmation for production automation.
- Agent skills vs. MCPAgent skills vs MCP explained: compare instructions, tool connections, security boundaries, versioning, and when to combine both for reliable AI agents.
- Production guardrails for AI agentsAI agent production guardrails for secrets, confirmations, audit logs, scoped access, structured errors, and safe autonomous deployment workflows.
- CI/CD with an agent tokenAI agent CI/CD with DOCKUP_TOKEN: authenticate without a browser, deploy with terminal-state waiting, protect secrets, and fail pipelines correctly.
- Security practices worth keepingProtect your applications with these essential security measures every developer should implement.
- What changes when an agent writes the codeExplore how AI-powered tools like Dockup Architect are revolutionizing the way we build and deploy applications.
Questions this page gets asked
Which agents does this work with?
Any agent that can run a shell command. Claude Code and Codex are the two it is tested against most, and there is a skill that installs the platform knowledge into either.
Can an agent delete something by accident?
Not without an explicit confirmation. Destructive operations refuse and return a recognisable code rather than a prompt, so the agent surfaces the decision instead of answering it.
Why JSON output rather than readable text?
Because human-formatted output changes wording between versions, and anything parsing it breaks silently. Every command accepts --json and returns a documented structure; the readable form stays for people.
Start on the free plan.
Ten dollars of infrastructure credit is already there, and no card is required.