Expose dashboard configuration as MCP-ready tools #12

New issue

Open

opened 2026-06-18 16:13:37 +02:00 by vince · 0 comments

vince commented

2026-06-18 16:13:37 +02:00

Owner

Parent: #1

Goal

Expose the dashboard configuration operations in a form that AI agents can discover and call through MCP or an equivalent tool protocol.

Scope

Use @modelcontextprotocol/sdk when this PR is reached, after the internal agent-safe configuration API is stable.
Define tool metadata, names, descriptions, input schemas, and output schemas for the core dashboard configuration operations.
Reuse TypeBox/JSON Schema definitions where practical so MCP tool schemas stay aligned with the dashboard model and internal API.
Reuse the internal agent-safe configuration API rather than duplicating mutation logic.
Add tool responses designed for agent loops: success, validation failure, preview diff, publish result, and rollback result.
Include examples for common tasks: add a service, add a telemetry card, connect a Prometheus query, rearrange a section, and preview before publish.
Add a local/dev transport or documented integration path for the first version.
Keep permissions scoped so tool calls cannot bypass validation or persistence history.

Acceptance criteria

Tool schemas are generated from or kept in sync with the dashboard model/API schemas.
An agent can preview and publish a simple dashboard edit through the tool surface.
Tool calls produce persisted revisions only through the validated write path.
Failure responses include enough structured detail for an agent to repair and retry.
Documentation explains the intended agent workflow.
MCP packaging does not become a prerequisite for the core dashboard runtime.

Out of scope

Natural-language chat UI.
Multi-tenant auth.
Hosted marketplace/plugin packaging.

Parent: #1 ## Goal Expose the dashboard configuration operations in a form that AI agents can discover and call through MCP or an equivalent tool protocol. ## Scope - Use `@modelcontextprotocol/sdk` when this PR is reached, after the internal agent-safe configuration API is stable. - Define tool metadata, names, descriptions, input schemas, and output schemas for the core dashboard configuration operations. - Reuse TypeBox/JSON Schema definitions where practical so MCP tool schemas stay aligned with the dashboard model and internal API. - Reuse the internal agent-safe configuration API rather than duplicating mutation logic. - Add tool responses designed for agent loops: success, validation failure, preview diff, publish result, and rollback result. - Include examples for common tasks: add a service, add a telemetry card, connect a Prometheus query, rearrange a section, and preview before publish. - Add a local/dev transport or documented integration path for the first version. - Keep permissions scoped so tool calls cannot bypass validation or persistence history. ## Acceptance criteria - Tool schemas are generated from or kept in sync with the dashboard model/API schemas. - An agent can preview and publish a simple dashboard edit through the tool surface. - Tool calls produce persisted revisions only through the validated write path. - Failure responses include enough structured detail for an agent to repair and retry. - Documentation explains the intended agent workflow. - MCP packaging does not become a prerequisite for the core dashboard runtime. ## Out of scope - Natural-language chat UI. - Multi-tenant auth. - Hosted marketplace/plugin packaging.