Skip to main content
Version: Cloud

Tools

The MCP server exposes tools in two groups with different behaviour:

You do not call these tools yourself. You describe what you want, and the AI agent selects and calls the right tool. The reference below explains what each one does so you can tell whether the agent is doing the right thing.

Agent registration and tracing​

These four tools implement the session flow. They are called automatically, so normally you do not need to think about them.

ToolWhat it does
list_agentsLists the agents registered in a workspace. Called at the very start of every session. If no workspace is configured it returns the available workspaces instead, so the user can choose one.
register_agentRegisters the agent for the current session and sets up tracing. Creates a registration if the name is new, and returns the existing one if it is not, so it is safe to call every session. Must be called before any traces are sent.
patch_agentUpdates fields on an already registered agent. Used only when the user explicitly asks to change an existing registration — never as part of the initial flow.
send_traceRecords one interaction as an audit trail entry. Called silently after every response once registration is complete, and surfaced to the user only if it fails.

register_agent takes the agent's name (which the agent must get from the user, never invent), plus optional description, source, endpoint, environment, model, and end-user details. patch_agent takes the same fields plus a deployment_id to identify the agent, and send_trace takes the conversation input and output. You never supply these yourself — the agent fills them in.

Use case controls​

These nineteen tools serve the compliance-officer workflow: understanding what a use case must demonstrate, reviewing its controls, and assembling evidence.

The group is manual and opt-in. The agent engages only when the user explicitly asks to work on use case controls or compliance evidence. It never starts this work on its own. See Skills over MCP for how the agent is guided.

Structure​

Deeploy organises compliance work in a hierarchy, which is worth understanding before reading the tool list:

  • A workspace contains use cases.
  • A use case has a lifecycle stage — exploration, development, validation, production, or retirement.
  • The stage determines which of the use case's controls are applicable. Controls outside the applicable set are still returned, marked as not applicable.
  • Controls come from one or more control frameworks active on the use case.
  • Each control carries evidence, may reference documentation (PDFs and documentation templates), and has checks — some automated, some requiring an attachment.

The typical flow is: list_use_cases → get_use_case → list_use_case_controls → get_use_case_control → add or link evidence.

Finding use cases and controls​

ToolWhat it does
list_use_casesLists the use cases in a workspace. The starting point for the workflow.
get_use_caseReturns one use case's lifecycle stage and control frameworks. Call before inspecting controls, so you know which stages are applicable.
list_use_case_controlsLists a use case's controls, marking each applicable or not for its stage.
get_use_case_controlReturns one control in full: its evidence, evidence linked from other controls, and its checks.

get_use_case_control accepts include_deployment_details to also fetch the use case's deployments, including each active version's model card, data card, and metadata. Those are an information source for writing evidence — reading them helps the agent write specific evidence, but they are not evidence themselves.

Changing a control's status​

ToolWhat it does
set_use_case_control_statusSets a control to completed or inProgress.
request_use_case_control_status_changeAsks the use case's owners to change the status (creates a task for them).

Only use case owners can set a status directly. If the caller is not an owner, the change is refused and the request route is used instead — this notifies the owners rather than failing silently.

Adding and editing evidence​

ToolWhat it does
add_control_evidenceAdds one evidence block to a control.
update_control_evidence_blockUpdates an existing evidence block. Only pass the fields that change.
delete_control_evidence_blockDeletes an evidence block. Irreversible — confirm before running.

add_control_evidence supports these evidence types:

evidence_typeUse forSupply
textMarkdown describing what exists and what it demonstrates. Right choice for evidence grounded in a deployment's model card, data card, or metadata.content
linkAn external resource.url
linked_documentAn existing imported PDF.imported_documentation_id
linked_templateAn existing use case documentation template.use_case_documentation_template_id

Image evidence is not supported over MCP — it requires a file upload, so add images in the Deeploy interface instead.

Evidence cannot be added to or edited on a control that is already completed. Change its status back to in-progress first, in the interface.

Every write tool accepts dry_run, which validates and returns the payload without creating anything. Agents are expected to show the proposed block and get approval before writing.

Linking evidence and check resources​

Evidence is written per control, but a control often needs evidence another control already owns. These tools create references rather than copies:

ToolWhat it does
link_control_evidenceReferences evidence owned by another control in the same use case.
list_control_evidence_linksLists evidence linked into a control from other controls.
unlink_control_evidenceRemoves an evidence link. The underlying evidence is untouched.

Controls also have checks, some of which require a document or template to be attached:

ToolWhat it does
link_control_check_resourceAttaches a resource to a linkable check.
list_control_check_linksLists the check links on a control.
unlink_control_check_resourceRemoves a check link, freeing the check to be linked again.

Only two kinds of check are linkable: document_linked (takes an imported document) and documentation_template_completed (takes a documentation template). Each check accepts one link — remove the existing one before attaching a different resource.

PDF (linked document) and template evidence cannot be linked between controls, and evidence already linked is rejected.

Documentation and templates​

Some controls rely on documentation templates that need to be filled in:

ToolWhat it does
list_use_case_documentsLists the PDFs available in a use case.
list_use_case_documentation_templatesLists the documentation templates attached to a use case.
get_documentation_template_questionsReturns a template's questions as a fillable answer map, with any saved answers.
fill_documentation_templateSaves answers using the indexes from get_documentation_template_questions.

The two template tools are used as a pair. get_documentation_template_questions returns the exact key structure the save expects, pre-populated with saved answers and null for unanswered ones. Answers must be markdown.

Saving merges, so a partial fill is fine — unanswered questions are left as they are. Template document content cannot be returned over MCP; to read a PDF, open it in the Deeploy interface.

Next​