Skills, AAS MCP, and Other MCP Tools

The short version is:

Skills are reusable SKILL.md playbooks that tell an AI assistant how to execute a workflow.
AAS MCP is the local, read-only interface through which an agent retrieves the verified AAS catalog without ranking, inspects evidence, and records its exact selected IDs in a proposed stack.
Other MCP tools connect an assistant to external systems such as APIs, databases, browsers, or hosted services.
The aas CLI validates the durable project profile and agent-selected stack, then previews exact lifecycle operations under human control.

These surfaces are complementary. AAS Core connects them around the approved aas-stack.json manifest.

What a skill does

A skill gives the model better instructions for a repeated task such as:

planning a feature;
reviewing code;
running a security audit;
writing a README;
debugging a failing test suite.

In practice, a skill improves the assistant's decision-making, structure, and process. It remains content, not an executable capability or a grant of authority.

Examples:

@brainstorming helps the model clarify requirements before implementation.
@lint-and-validate helps the model run appropriate quality checks before claiming success.

What AAS MCP does

AAS MCP is part of this repository's product, not an unrelated external integration. Codex or Claude uses it to call the same deterministic AAS Core that powers the CLI projections.

It exposes exactly:

search_skills;
get_skill;
compose_stack;
inspect_stack;
diff_stack;
the aas://skills/{id} resource template.

The agent inspects the project using its normal local capabilities, searches and reads the complete catalog, and chooses the exact skills. AAS MCP does not scan the repository itself, and Core does not rank or exclude candidates. Catalog metadata is informational only.

AAS MCP is local stdio, process-per-session, read-only, offline-capable, and contains no model credentials or telemetry. Its tool calls do not install or remove skills, apply a stack, update catalogs, or edit host configuration.

What other MCP tools do

An external MCP tool may give the model a capability it would not otherwise have, such as:

reading from a database;
calling GitHub APIs;
fetching documentation from a service;
creating calendar events;
querying or changing another system.

Those tools expand what the assistant can do in the world. Their permissions, network behavior, and write boundaries depend on each server. They are different from the deliberately narrow AAS MCP catalog/composition boundary.

What the CLI does

The aas CLI is the explicit operational interface for:

catalog status and updates;
MCP configuration;
validation and composition of an agent-selected stack;
manifest validation;
plan creation and read-only diagnostics.

The durable artifact is aas-stack.json, which pins catalog identity, targets, the project profile, and exact skill IDs. The agent proposes it; the user reviews it. It has no Core selection policy. stack validate checks it, and stack plan creates an immutable preview bound to the observed managed state and exact operations.

stack apply and stack recover exist only for controlled preview development. They are experimental, disabled by default, and are not supported or certified preview safety claims.

The easiest mental model

Skills provide the operating guidance.
AAS MCP gives the agent complete catalog access and records the selection the agent makes.
aas-stack.json records what the user approved.
The CLI validates and previews the lifecycle.
Other MCP servers provide access to outside systems.

Which path should you start with?

Start with AAS Core when:

you use Codex or Claude with local MCP support;
you want the agent to compose a small stack from project evidence;
you need complete catalog access and a reproducible manifest of the agent's exact selection;
you want a reviewable plan before any skill changes.

Start with a direct skill or specialized plugin when:

the host does not yet have a native AAS MCP configuration adapter;
you already know the exact skill or fixed domain pack you want;
you prefer manual invocation without a managed stack lifecycle.

Add other MCP tools when the work needs live access to APIs, services, databases, or hosted platforms. A stack can contain skills that guide those integrations without granting the external permissions itself.

Preview limits

AAS Core helps Codex and Claude search and read the complete local catalog, choose exact skill IDs, and preserve them in a reproducible stack. Metadata may be reported as unknown, but it is informational and never blocks selection or use. Apply and recovery remain experimental.