Holon

Create and run AI assistants on your own computer or server. Assign tasks, track progress, and let them follow up. Open source for individuals and teams.

Skills and tools

A skill is a reusable workflow rooted at a SKILL.md file. A tool is an operation the runtime makes available, such as reading a file, running a command, or searching memory. The agent reads a skill to learn a procedure, then carries it out with available tools.

For example, a code-review skill can require evidence for each finding. The agent may use a command tool to inspect a diff and run tests. Installing that skill does not grant shell access, repository credentials, or permission to submit a review.

How a skill becomes useful

Skill management and skill use happen at different stages:

  1. The Skill Library records skills available for managed use.
  2. Enabling a library skill makes it available to a particular agent.
  3. Discovery exposes a catalog entry with its name, description, and path.
  4. When a task matches, the agent opens SKILL.md and follows the workflow.

The catalog can also include skills discovered in the active workspace. A visible entry means the agent can find the skill; it does not mean the full instructions are already loaded or that the workflow has run.

This distinction helps diagnose problems. If a skill is missing, inspect discovery and enablement. If the workflow was skipped, inspect whether the agent read the skill. If a command failed, check the tool, credentials, and environment rather than reinstalling the workflow by default.

Guidance and authority

AGENTS.md defines standing responsibilities and project conventions. Skills explain repeatable tasks within that scope. An Agent Template can combine a role contract with skill dependencies so new agents start with the expected workflows.

A skill may describe a step that the current task does not permit. For a read-only review, a publication step remains out of scope. Skill text cannot enable a tool that runtime policy withholds or turn external input into an operator instruction. Scripts bundled with a skill are code to inspect and run under the same execution limits as other commands.

Updates and shared sources

Library management and per-agent enablement are separate. Removing or updating shared library material can affect more than the agent currently in view. Disabling a skill for one agent changes that agent's managed enablement.

Linked skills reflect their source files. A managed copy has its own update path. Choose deliberately when you need a shared evolving workflow or an independently managed copy, and verify the effective skill list after a change. Installing a local link does not update the source repository.

What to expect from tools

Tool availability comes from runtime capabilities and policy. A provider or execution backend may impose additional requirements: image inspection needs a usable vision route, and a command needs an execution environment. A workflow can document these requirements but cannot satisfy them by being installed.

Tool receipts provide evidence of what happened. Reading a skill is not proof that a test passed; an accepted request is not proof that its work completed. Check the result appropriate to the operation.

Next steps