AI agents

JEKYO is built to be operated by AI coding agents. The CLI is the agent API.

Teach your agent once

jekyo skill install --global

This installs an instruction pack user-level for Claude Code (~/.claude/skills) and Codex (~/.codex/AGENTS.md). From then on, in any session:

/jekyo deploy this app
/jekyo why is the api service crashlooping?

The pack teaches the agent the complete DSL (with "keys not listed do not exist" strictness, so no hallucinated config), and the operating discipline: check the context, always jekyo render before jekyo up, verify with jekyo ps and jekyo status.

Project-scoped installs are available too (jekyo skill install in a repo detects .claude/, AGENTS.md, and .cursor/), and any other agent can bootstrap itself:

jekyo skill show     # the full reference on stdout
jekyo schema         # JSON Schema for jekyo.yaml

What it looks like

A real session, after jekyo skill install --global:

you>   /jekyo deploy this app to my server

agent> Checking the target: jekyo context show    (server: prod, 1.2.3.4)
agent> Validating: jekyo render                   (2 services, 1 volume, OK)
agent> Deploying: jekyo up
       App shop deployed (revision 4, 2 services)
         api: https://api.shop.example.com
agent> Watching rollout: jekyo ps shop            (2/2 Running)
       Done. Your app is live at https://api.shop.example.com

you>   /jekyo the api feels slow since the deploy

agent> jekyo logs shop/api --since 15m
       ... connection pool exhausted ...
agent> The api logs show pool exhaustion after revision 4.
       Want me to roll back? jekyo rollback shop v3 is one command.

Every step is a plain CLI call you could have typed yourself. Nothing hidden, everything auditable.

Why the CLI, not a plugin protocol

Agents with shell access already run CLIs natively. JEKYO leans into it:

• Structured output: -o json on read commands (ls, ps, history, ...), and jekyo top --json for a full resource snapshot (per-pod CPU, memory, restarts, node capacity) in one call.
• Self-checking: jekyo render shows exactly what a jekyo.yaml becomes; validation errors are precise and actionable.
• Scoped credentials: an agent operates through a context, not raw SSH keys and kubeconfigs scattered around.
• Safe reads: logs, status, ps let an agent investigate production without any ability to change it.

The result: "deploy this, watch the rollout, tail the logs, roll back if it breaks" is a conversation with your agent, and every step of it is a command you could have typed yourself.