Teach your agent

Agents learn Zeroko from a skill: a Markdown file that explains how to write a service, deploy it, use each binding and read the logs. Every panel serves its own, filled in with where the API is and its limits (heap, CPU time, disk quota), so what the agent reads matches where it deploys.

What the agent needs

• The skill: on Zeroko Cloud, https://zeroko.space/workspaces/<workspace>/skill.md (the workspace's Services page copies it), for its members signed in and for agents with one of its tokens; self-hosted, https://panel.example.com/skill.md. Zeroko Cloud's skill says its version, and asks the agent to fetch it again before each use and follow the new one if it changed.
• A token: make one in the panel, under Tokens. Tokens start with zk_, can expire or last forever, and can be revoked at any time.

The agent calls the API with Authorization: Bearer zk_...: on Zeroko Cloud at the workspace's own address, https://<workspace>.zeroko.space/api; self-hosted at https://panel.example.com/api, which the panel passes to its runner. Either way the runner stays private.

Claude

In Claude Code, add the skill to your skills folder (it keeps itself up to date: it checks its address before each use), or ask Claude to read it at the start of a session:

Read https://panel.example.com/skill.md. Use the token in ZEROKO_TOKEN to
deploy services to Zeroko.

On claude.ai, add the skill to your skills and the token to the project's instructions, or let Claude keep it in its memory, not in the skill: a skill's text can be shared.

MCP

The same API is an MCP server, for agents that speak the Model Context Protocol: on Zeroko Cloud at https://<workspace>.zeroko.space/mcp, self-hosted at https://panel.example.com/mcp, with the same token. Its tools deploy, inspect, roll back, stop and delete services, read their logs and metrics, set secrets and vars, query their databases and run their crons. Each one is an API call made with the agent's token, so it can do what the token can, and no more.

In Claude Code:

claude mcp add --transport http zeroko https://acme.zeroko.space/mcp \
  --header "Authorization: Bearer zk_..."

Any client that takes a remote MCP server over HTTP, with a header, works alike. The tools say what to do, not how to write a service: the skill still teaches that, and goes well with them.

Other agents

Any agent that can make HTTP requests can use Zeroko: give it the skill's text in its instructions and the token in its environment.

What a good session looks like

1. The agent writes the service, in a few modules.
2. It deploys. On 422 it reads deployment.error, fixes the code and deploys again; the last good version keeps serving meanwhile.
3. It requests the service's address and reads the logs to check it works.
4. It sets the secrets the service needs; they apply at once.
5. On an error at runtime it reads the logs, fixes and deploys again, or rolls back to an earlier version.