Skip to content

MCP Tools

The Tools tab gives you granular control over every tool your MCP server exposes to AI clients. Toggle tools on or off, organize them by group, edit descriptions, set behavioral annotations, define system rules, and configure cache policies — all from a single interface.

Most MCP platforms treat tools as static definitions: whatever your OpenAPI spec declares is what the AI model sees. Vinkius Cloud introduces server-side tool governance — you control which tools are active, how the AI should interpret each one, and what behavioral constraints apply, without modifying your upstream API.


Tool list

Tools are displayed in a searchable list with per-tool toggles. The view supports two layout modes:

Grouped view (default)

Tools are organized by their OpenAPI tag. Each group has a collapsible header with a group-level toggle:

MCP TOOLS
+ Add Tool
USERS4
ALL_ON
GETlist_users/api/v1/users
POSTcreate_user/api/v1/users
DELdelete_user/api/v1/users/{id}
5/6 ENABLED2 GROUPS

Flat view

When the server's tool exposition is set to flat, the same tools are displayed in a single list without group headers — useful for servers with no OpenAPI tags or few tools.

Tool columns

ColumnDescription
ToggleEnable or disable the tool. Disabled tools are excluded from tools/list responses and cannot be called
MethodHTTP verb (GET, POST, PUT, PATCH, DEL) with color coding
NameTool name as exposed to AI clients. Click to open the editor
PathThe upstream API endpoint path

The search input filters tools by name, endpoint path, or description. The filter applies across all groups in real time.

Group toggle

Click the group-level toggle to enable or disable all tools in a tag group with a single action. The toggle shows ALL_ON or PARTIAL status.


Tool editor

Click any tool name to open the slide-out editor panel. The editor provides full control over how the AI model sees and interacts with the tool:

EDIT TOOL
list_users
x
ENDPOINT
GET/api/v1/users
Original: Returns a paginated list of all users...
SYSTEM RULES
Never return more than 20 results
Always include pagination metadata
DISPLAY TITLE
List Users
CUSTOM DESCRIPTIONReset to Original
Retrieves a paginated list of users. Supports filtering by role, department, and status.
This overrides the original OpenAPI description
LLM ANNOTATIONS
Read Only
Only retrieves data, no side effects
Destructive
May perform irreversible changes
Idempotent
Repeated calls have no additional effect
Return Direct
Output goes directly to the user
CACHE CONTROL
immutablev
Delete
CancelSave

Editor sections

System Rules

Behavioral constraints injected into the tool's context. One rule per line. The AI model receives these rules alongside the tool definition, guiding how it constructs requests and interprets responses.

Examples:

  • Never return more than 20 results
  • Always include pagination metadata
  • Omit internal fields: _id, __v

Identity overrides

FieldPurpose
Display TitleOverride the tool name shown to AI clients (e.g., List Users instead of listUsers_get_api_v1_users)
Custom DescriptionReplace the auto-generated description with a human-written one. A Reset to Original button restores the OpenAPI description

Why override descriptions

Auto-generated descriptions from OpenAPI specs are often technical and incomplete. A well-written description directly improves how accurately the AI model selects and uses the tool. This is one of the highest-impact changes you can make per tool.

Annotations

MCP tool annotations are behavioral hints that tell AI clients how a tool operates. Toggle each annotation on or off:

AnnotationEffect
Read OnlyIndicates the tool only retrieves or computes information without modifying data
DestructiveWarns that the tool may perform irreversible changes (deletes, overwrites)
IdempotentSignals that repeated calls with the same arguments produce no additional effect
Return DirectTells the client to return the tool's output directly to the user without further processing

AI clients use these annotations to make safety decisions — for example, automatically approving read-only calls while requiring confirmation for destructive operations.

Cache Control

Configure how AI clients should cache this tool's responses:

OptionBehavior
NoneNo cache directive — clients decide
no-storeExplicitly prevents caching. Use for real-time data (balances, live status)
immutableMarks the response as permanently cacheable. Use for static data (schemas, enums, configurations)

Adding a manual tool

Click Add Tool to define a tool manually — useful for endpoints not covered by your OpenAPI spec import. Required fields:

FieldValidation
NameRequired, max 255 characters
HTTP MethodGET, POST, PUT, PATCH, or DELETE
Endpoint PathRequired, max 500 characters
DescriptionOptional, max 1,000 characters
TagOptional group name, max 100 characters

Manual tools are immediately available to connected AI clients after creation.


Deleting a tool

In the tool editor, click Delete at the bottom. A confirmation modal requires you to type the exact tool name before deletion proceeds. This prevents accidental removal of critical tools.

After deletion, all active SSE connections are notified via mcp:invalidate so connected clients refresh their tool list immediately.


Agent Skills

Below the tools list, the Tools tab also displays Agent Skills — markdown-based instructions that extend your MCP server's capabilities beyond API tools.

Skills are distinct from tools: while tools map to HTTP endpoints, skills provide contextual knowledge, decision frameworks, and behavioral guidelines that guide the AI model during conversations.

Skill list

AGENT SKILLS
+ Add Skill
OnNameDescriptionActions
code-review-guidelinesStandards and patterns for reviewing pull requests 🗑
deployment-runbookStep-by-step production deployment procedures 🗑

Each skill has a violet toggle (distinct from the cyan tool toggles) to activate or deactivate it. Disabled skills are excluded from the tool listing sent to AI clients.

Skill Studio

Click a skill name or the edit icon to open Skill Studio — a full-screen editor designed for writing and editing skill documents:

🧠Skill Studio/code-review-guidelinesACTIVE
CancelSave
SKILL
NAME
code-review-guidelines
Max 64 characters
DESCRIPTION
Standards and patterns for reviewing pull requests in our codebase
62 / 1024
INFO
Lines47
Characters1,842
SKILL FILES
Drag and drop auxiliary files — scripts, references, assets.
SKILL BODY--Markdown supported
1 # Code Review Standards
2
3 When reviewing code, follow these principles:
4
5 - Check for proper error handling
6 - Verify input validation
7 - Ensure test coverage for new logic
Ready
Ln 7, Col 138 words1,842 charsMarkdown

Skill Studio features

AreaDetails
Top barBrain icon, name breadcrumb, ACTIVE badge (when skill is enabled), Cancel and Save buttons
Left sidebarName (max 64 characters), Description (max 1,024 characters with counter), document stats (lines, characters)
Skill FilesDrag-and-drop area for auxiliary files — scripts, references, assets. Available after the skill is created. Storage limits are plan-dependent
EditorFull Monaco editor with syntax highlighting, line numbers, and minimap. Supports markdown
Status barReady/Saving indicator, line/column position, word count, character count
FrontmatterIf you paste content with YAML frontmatter (name/description fields), the editor auto-extracts them into the sidebar fields

Skills vs Tools

Tools map 1:1 to HTTP endpoints. They have methods, paths, and generate API calls. Skills are knowledge documents. They provide context — guidelines, runbooks, decision frameworks — that the AI model can reference during conversations. Use skills for everything that is not a direct API call.


Real-time propagation

Tool and skill changes propagate to connected AI clients automatically through two mechanisms:

EventTriggerEffect
mcp:tools-changedToggle, edit annotations, update descriptionConnected clients receive a notifications/tools/list_changed event and refresh their tool list
mcp:invalidateAdd or delete a toolActive SSE connections are fully re-established with the updated tool set

This means you can adjust tool availability, descriptions, and annotations in production without requiring clients to reconnect or restart.


Next steps


Frequently Asked Questions

Can I disable a tool without deleting it?

Yes. Each tool has an individual toggle. When disabled, the tool is excluded from tools/list responses and cannot be called by AI clients. The configuration, description, annotations, and system rules are preserved. Re-enable it with a single click.

What are MCP tool annotations and why do they matter?

Annotations are behavioral hints that tell AI clients how a tool operates: Read Only (no side effects), Destructive (irreversible changes), Idempotent (safe to retry), and Return Direct (output goes straight to the user). AI clients use these annotations to make safety decisions — automatically approving read-only calls while requiring confirmation for destructive operations.

Can I customize tool descriptions without modifying my API?

Yes. The tool editor lets you override the auto-generated OpenAPI description with a human-written one. A well-written description directly improves how accurately the AI model selects and uses the tool. You can always reset to the original description with one click.

What are Agent Skills and how are they different from tools?

Tools map 1:1 to HTTP endpoints and generate API calls. Skills are markdown-based knowledge documents — guidelines, runbooks, and decision frameworks — that the AI model can reference during conversations. Skills extend your MCP server's capabilities beyond API tools without requiring custom code.

Do tool changes propagate to connected AI clients automatically?

Yes. When you toggle, edit annotations, or update descriptions, connected clients receive a notifications/tools/list_changed event and refresh their tool list. When you add or delete a tool, active SSE connections are fully re-established with the updated tool set. No client restart required.

What cache control options are available per tool?

Three options: None (clients decide), no-store (prevents caching — use for real-time data like balances or live status), and immutable (permanently cacheable — use for static data like schemas, enums, or configurations).