01
Declared scope
Every tool, resource and prompt names the scope it needs and the evidence it returns.
Agents, MCP & capabilities · Concept
AgentPlugin treats integrations as declared interfaces, not as adapters written per customer.
Every row is a specification requirement. None is claimed as live.
| Surface | What it exposes | Status |
|---|---|---|
| REST / OpenAPI | Resource API for the product's core objects, described by an OpenAPI document. | Specified |
| GraphQL | Optional typed query surface for composite reads. | Specified |
| MCP tools | Tool definitions for agent invocation, scoped per capability. | Specified |
| MCP resources | Read-only resources exposed to connected agents. | Specified |
| MCP prompts | Reusable prompt templates with declared inputs. | Specified |
| MCP Tasks | Long-running work represented as a task with status and evidence. | Specified |
| MCP Apps | Embedded application surfaces that call back into the product. | Specified |
| Skills over MCP | Packaged skills discoverable and callable through MCP. | Specified |
| WebMCP | Browser-side MCP bridge for local, in-page capabilities. | Specified |
| A2A Agent Card | Published agent card for cross-agent discovery. | Specified |
| ACP | Agent communication profile for delegated work. | Specified |
| AG-UI / A2UI | Interface contracts for agent-driven user surfaces. | Specified |
| Webhooks | Signed event delivery with retry and replay. | Specified |
| CLI | Scriptable local entry point with the same permission model. | Specified |
| SDKs | Typed client libraries generated from the API description. | Specified |
What an integration has to prove before it is listed.
Every tool, resource and prompt names the scope it needs and the evidence it returns.
The same policy applies to a human caller, an SDK and an agent; no surface is a bypass.
An invocation can be reconstructed from its inputs, policy revision and output record.