Building Kavor #01: Agents Have Invisible Superpowers

[digows]

Coding agents already have superpowers. The awkward part is that many of those powers are invisible.

A provider may discover instructions from the filesystem, load a plugin, connect to an MCP server, inherit permissions, or receive tools from its runtime. The user sees the result, but rarely gets a clear operational model of what made it possible.

What we are exploring

Kavor already makes agents, work, Connections, and Guardrails visible on a Canvas. The next question is whether a CodingAgent Node can also explain:

1. What can this agent do?
2. Why is that capability available?
3. What can it access or change?
4. Is it ready to use?
5. How do I disable or remove it?

A distinction worth protecting

Our current working model separates:

• Skill: instructions, knowledge, references, and a repeatable workflow.
• Connector: executable access to an external system through MCP, an API, a browser, or a native application.
• Capability Pack: a possible installable unit that combines Skills, Connectors, compatibility requirements, and presentation metadata.

A Skill can teach an agent how to diagnose an n8n workflow. It cannot grant n8n access by itself. That requires a Connector, authentication, and an authorization boundary.

The first real test

We do not want to design a generic marketplace in isolation. The first Capability Pack should be narrow enough to expose flaws in the model.

The leading candidates are:

• GitLab, because it is close to the software-development workflow and tests read/write authorization;
• n8n, because it combines API operations, workflow knowledge, and clear visual outcomes.

Browser control and native applications are intentionally later. They introduce separate runtime problems that would hide the capability-model questions we need to answer first.

Your turn

Think about the most capable coding-agent setup you use today:

• Which capability would another developer fail to discover?
• Which access would you be uncomfortable granting silently?
• Should Skills and executable Connectors appear together or separately?
• Which real integration would best test whether this model is honest?

No formal proposal required. A concrete workflow or sharp counterexample is enough.

Read the complete, versioned track: Visible Agent Capabilities.