# Using the vault with your agent

A user can supply this vault URL and a task without installing an individual skill. The agent still needs an access method and the capabilities required by that task. Opening a page alone does not authorize execution or install its instructions.

## Access methods

| Host access | How to use the same content | Boundary |
| --- | --- | --- |
| HTTP or web fetch | Read README, the catalog, one pack and its linked instruction/reference URLs | No JavaScript or MCP required |
| Browser control | Browse the folder explorer, open the skill and needed references | The host applies the retrieved text |
| Connected MCP | Search, describe, load and read resources at one exact release | Client connection is required, individual skill installation is not |
| Browser WebMCP | Use the page's read-only tools when the host exposes them | Browser and agent support must be observed, not assumed |

## Translating installed conventions

- A slash command such as `/blog outline` names a workflow. On this website, select `blog-outline` and read its original SKILL.md. It is not a command that every agent understands automatically.
- A repository path names a source file. Resolve it through the selected pack's declared resource path and pinned URL. Do not substitute an unrelated same-named local file.
- A linked agent Markdown file describes a role. Read it as supporting instructions. It does not spawn another agent; delegation depends on the host and the user's task.
- A local output path describes an intended artifact. Use the user's actual workspace or return a reviewable document when file writing is unavailable.
- A script, API, account, image generator or external service remains a real dependency. Reading its name cannot supply that capability. Report an unavailable step; do not claim its checks passed.
- A reference to another skill means discover and describe that skill separately at the same release. Do not pretend its instructions were read merely because its name appears.

## Execution status

Every imported pack is reference guidance. The descriptor records requirements, effects and known runtime gaps. Its preflight compares declarations with supplied host facts; it does not test a tool or grant permission.

- **Supported for this task:** the host has the required inputs and capabilities, and can perform the requested checks. Establish this for the actual task, not for the entire skill collection.
- **Conditional:** useful work is possible but a dependency or evidence source is missing. Label provisional output and state the omitted checks.
- **Unavailable:** a requested result depends on a capability the host cannot provide. Stop that step and explain the specific requirement.

Writing from supplied facts, inspecting text and planning can often proceed with text access. Live research needs research tools. Image/audio generation and CMS operations need their respective services. Installed helper scripts and their deterministic gates are not replaced by a model's judgment.

## A first task

Use this vault to outline an article for adult beginners about organizing a home bookshelf. Read the relevant skill and references. Use only supplied facts; label the outline provisional if live research is unavailable. Do not invent citations, install software or publish anything.

Start with [README](/mcp/README.md), [AGENTS](/mcp/AGENTS.md), [source scope](/mcp/SOURCE.md) and [trust boundaries](/mcp/TRUST.md).