---
name: mcp-integration-development
description: "Design and verify bounded MCP tools and agent integrations from explicit contracts and authorized capabilities. Use for mcp integration development tasks; select only the requested methods."
metadata:
  version: "1.0.0"
---

# MCP Integration Development

Design and verify bounded MCP tools and agent integrations from explicit contracts and authorized capabilities.

## OLLIE Runtime integration

Use relevant methods under the active OLLIE Operator. Read live Workspace Context and tenant bindings for client facts, account identity, goals, permissions and required inputs. Consume ask-ollie-authority-core for authority, outloop-access-fallback for access and recovery, email-lifecycle-core for communication, and the bootstrapper Runtime Contract for execution ownership and completion. This skill introduces no scheduler, independent permission policy or competing Runtime owner.

Discover the active Workspace’s approved capabilities before choosing API or browser access. Use Outloop for external service actions and the host’s supported tools for local artifacts. A product name in this package is not an access grant. Never import private account configuration or follow a source method that bypasses host restrictions. Apply the current task’s authorization; a library update does not grant permission to send, publish or activate a campaign.

## Required inputs

- Integration goal and host
- Official protocol and provider documentation
- Tool input/output contracts
- Authorized development workspace and synthetic fixtures

Use for MCP server/tool design, agent-facing integration contracts and related plugin packaging. Read [integration contracts](references/contracts.md). Use host-native development tools already authorized for the task.

## Define the boundary

Specify the desired tool outcome, inputs, outputs, errors and side effects. Keep read operations separate from writes and annotate mutation semantics accurately. Tool descriptions should guide correct selection without suggesting broader authority. In an Outloop workspace, external access remains routed through approved capabilities; a new MCP tool must not become a credential bypass.

## Implement against current contracts

Consult the [official MCP documentation](https://modelcontextprotocol.io/docs) and the intended host’s current plugin requirements. Verify negotiated capabilities, schema validation, pagination, cancellation and transport behavior rather than relying on a remembered SDK version. Do not install unrequested connectors or broaden resource access.

## Test and package

Use synthetic provider responses for normal, malformed, empty, partial, paginated, unauthorized, rate-limited and uncertain-write cases. Verify tool names and descriptions remain discoverable, output stays bounded and sensitive data is redacted. Separate local protocol success from real provider proof.

Package with the target host’s supported manifest and complete references. For a tool-backed UI or downloadable plugin, verify its actual install route and permission surface. Do not claim every agent host supports the same importer.

## Completion

Verify findings against source evidence and read back every resulting artifact or authorized change. Distinguish a saved draft, queued job, active workflow and completed delivery. Return the requested outcome, evidence, output location and precise outstanding dependencies. Complete independent work when one dependency is unavailable. Package validation and synthetic fixtures do not establish real host import or provider execution.

Read [evaluation scenarios](references/evaluation.md) when testing this skill.