Skills
Procedural knowledge for agents, published separately from tools and carrying no signing authority. Planned, not built — scoped by the agent-surfaces decision, which ships in the repository.
Knowing how to compose a direct mint safely is not the same capability as being able to submit one. Skills are the first thing; they are never the second.
Planned. This surface is scoped and not yet built. The decision that scoped
it is .thoughts/decisions/2026-08-03-agent-facing-surfaces.md; decisions ship
in the repository, so you can read the reasoning before any of this exists.
There is no skill to install today.
What it will be#
A skill carries procedural knowledge — how to sequence an operation, what to check before it, what the recovery paths are. It is never a transaction tool and never implies signing authority, and it is published separately from the tools so a host can adopt the knowledge surface without adopting any action surface.
Each skill declares its name, description and license, and states explicitly that it cannot sign. Skills will be installable through the CLI and through a plugin marketplace entry, from one source of truth with no duplicated content between the two.
Why they are separate#
The three deliveries divide cleanly, and the division is the point:
- MCP connects an agent to this system's actions.
- Skills carry procedural knowledge about using it safely.
- CLI gives humans and shell agents the same surface deterministically.
A skill is also far cheaper than a server: an MCP server injects every tool schema into the host's context at startup, where a skill is read when it is needed.
What exists today#
The documentation you are reading, and the headless operation contract it describes. A skill will encode the same procedures — see the quickstart and mock mode for the lifecycle those procedures are about.