Veasel Code targets the released Agent Plugins v1.0.0 specification. A compatible host must understand packages, Skills, and MCP configuration, then add its own clear trust and permission controls.
◌
In progress · no compatibility claim yet
The V code currently validates local manifests, discovers Agent Skills, and parses MCP server configuration. It does not install or run plugins, connect MCP servers, activate Skills in a model turn, or ask for execution approval.
What the format brings
Plugin package
A portable directory starts with a versioned plugin.json manifest.
✳
Agent Skills
Discover skill descriptions first; load each instruction body only when used.
⌘
MCP servers
Declare stdio or Streamable HTTP connections in a package’s mcp.json.
What Veasel must add as the host
A managed plugin registry with source, version, enabled state, and persistent per-install data.
Isolated failure handling for malformed packages, unsupported components, and disconnected servers.
Explicit trust decisions and exact tool approvals; plugin metadata cannot grant itself authority.
Bounded MCP transports, cancellation, shutdown, and output limits.
Skills and MCP tools integrated into the model loop and reviewed in the TUI.
Specification-derived fixtures and a published compatibility report before claiming conformance.
An allowed-tools hint in a Skill is advisory input. It must never override user policy or Veasel’s permission checks.