Skip to main content
Version: 1.67

Skills over MCP

Skills are instruction documents that tell an AI agent how to carry out a workflow — as opposed to tools, which are the individual actions the agent can take. Delivering a skill gives the agent the full procedure in one step, so it runs the workflow consistently instead of improvising each time.

The Deeploy MCP server provides two skills, matching the two tool groups:

SkillPurposeWhen it runs
deeploy-autoregisterRegisters the session and traces every interaction.Automatically, at the start of every session.
deeploy-use-case-controlsGuides control review and evidence gathering.Only when a user explicitly asks about controls or compliance.

The split is deliberate, and the two skills behave very differently. Registration and tracing apply to everyone, so that skill is designed to fire automatically and unconditionally. Control work applies only to compliance officers, so its skill is designed to stay dormant until asked for.

How skills are delivered​

The skills are served by the MCP server over the Skills extension, and their content is also readable as MCP resources. Your platform needs to support sending skills for them to take effect; where it does not, the agent falls back to the MCP server's own instructions and the flow is less reliable.

Practically, this means you do not paste skill text into your agent configuration. The server provides it, and you add the system instruction that tells the agent to use the Deeploy server at the start of every conversation.

deeploy-autoregister​

Transforms the session flow into an explicit, ordered procedure so that registration and tracing happen reliably, even when the user's first message has nothing to do with Deeploy.

What it does:

  1. Calls list_agents first — always, before anything else — and presents the existing registrations.
  2. Asks the user for a registration name, then calls register_agent with the environment, model, and user identity.
  3. After registration, traces every subsequent interaction by silently calling send_trace.

Why it is written so forcefully. An agent's default behaviour is to answer the user's question and move on. Registration would therefore be skipped whenever the opening message looked unrelated — which is precisely when the audit trail matters. The skill is worded to override that instinct, which is why its description insists it runs before anything else and does not depend on what the user said.

Deliberate constraints the skill enforces:

  • The registration name must come from the user. The agent never invents or suggests one.
  • list_agents is always the first call, before asking for a name.
  • register_agent is used for initial registration — never patch_agent, even when the name already appeared in the agent list.
  • Traces are sent silently. They are surfaced only on failure.
  • Typing skip aborts registration and tracing for the session, and the agent proceeds with the user's request.

The result is that every session produces an audit trail without anyone remembering to set one up.

deeploy-use-case-controls​

Guides the agent through the use case controls workflow: understanding which controls apply, judging whether the existing evidence is genuinely sufficient, and assembling or linking what is missing.

What it covers:

  • Reading a use case's lifecycle stage and applicable control frameworks.
  • Reviewing controls and the evidence attached to them.
  • Adding evidence, or referencing evidence another control already owns.
  • Attaching documents and templates to checks that require one.
  • Filling in documentation templates.

It is strictly opt-in. The skill's instructions forbid the agent from engaging unless the user clearly asks about controls or compliance evidence. Trigger phrases include "help me fill in the controls", "what evidence is missing for this control", and "can our model card count as evidence for this control". If the user merely registers an agent or asks a general Deeploy question, the skill stays dormant.

Honesty over completeness. The most important instruction here is that the agent must not treat filled-in fields as compliance. Its job is to assess critically whether the evidence genuinely demonstrates the control — and to say so when it does not. A control that has been filled in is not automatically a control that has been satisfied.

Writing is proposed, not silent. Before creating or changing anything, the agent shows what it intends to write and gets approval. The write tools also support a dry-run mode that returns the exact payload without saving.