Turn a service or local capability into a traceable MCP server design with primitive decisions, tool contracts, authorization boundaries, tests, and release gates.
An MCP server can pass a happy-path demo while accepting the wrong token audience, mixing tenants, labeling a mutation as read-only, or returning enough data to swamp the model context. This skill keeps those decisions visible before they become production behavior.
It turns user jobs and service evidence into a server contract. It decides what belongs as a tool, resource, prompt, task, sampling request, or elicitation. Then it defines schemas, side effects, pagination, errors, scopes, tenant binding, confirmations, compatibility, negative-path tests, agent evaluations, and release gates. Protocol facts, service facts, assumptions, and proposed checks stay separate.
Use it for a new API integration, a local stdio server, a remote multi-tenant server, or a pre-release review. The package includes an MCP decision rubric, human and JSON report formats, a canonical plan schema, and a read-only checker for references, annotations, authorization safeguards, test coverage, gates, and verdict consistency. It does not inspect an API you have not supplied, run untrusted code, deploy a server, or prove security and client compatibility without executed tests.
Input
Design a Streamable HTTP MCP server for our issue API. Agents need to search projects, read issues, create comments, and close issues. The API uses tenant-scoped OAuth tokens. Claude Desktop and Codex are the first clients. Closing is destructive, comments notify watchers, and search can return 50,000 rows. Plan only. Do not invent endpoints that are not in the supplied OpenAPI file.
Output
An evidence register; capability-to-primitive decisions; bounded search, issue-read, comment-create, and issue-close contracts; schemas, pagination, annotations, scopes, tenant and audience binding, downstream-token separation, confirmations, and error classes; client and protocol compatibility rows; positive, denial, isolation, pagination, retry, and destructive-path tests; release gates; and a verdict limited to the supplied API evidence.
Provide the user jobs, service or local capability, API or code evidence, target MCP clients, deployment and transport, authentication model, tenants, sensitive data, SDK and runtime constraints, and the requested mode: plan, review, or implement. The skill returns a bounded server contract, trust boundary, compatibility matrix, implementation sequence, tests, release gates, limitations, and one next action.
API specification, SDK or source code, example payloads, local command contract, or other material that proves what the server can expose.
Local or remote deployment, transport, users, tenants, sensitive data, authentication and downstream authorization model, and owners for unresolved security decisions.
Target clients, MCP specification date, SDK and runtime version, and any required optional capabilities. Unknown targets can be recorded as blockers or planned checks.
Optional. Required only to run the bundled read-only canonical plan checker.
SKILL.md; MCP primitive, contract, security, compatibility, verification, and verdict rubric; human and canonical JSON report formats; MCP server plan JSON schema; read-only plan checker; Agent Skills interface metadata
No reviews yet.
Input
Plan and then implement an MCP server in this TypeScript repo for reading sanitized build logs, listing known services, and restarting a local dev service. It runs over stdio. Logs may contain secrets, service names come from a checked-in config, and restart must never accept an arbitrary shell command. Follow the repository instructions and show the change boundary before editing.
Output
A local-process trust boundary; resource-versus-tool decisions; path, size, redaction, service allowlist, stdout/stderr, mutation approval, timeout, and idempotency rules; exact proposed files and dependencies; read, invalid-service, injection, redaction, restart, timeout, and shutdown tests; executed verification when authorized; and no remote OAuth ceremony presented as relevant to stdio.
Input
Review this MCP design and return the canonical JSON record. One tool named execute accepts an endpoint and arbitrary JSON. It forwards the client's bearer token to the CRM, can delete contacts, sets readOnlyHint true, returns complete customer records, and has no tenant test. The remote server claims support for every MCP client but names no protocol or SDK version. No changes are approved.
Output
One schema-conforming JSON record that separates observed facts from unknowns; rejects the mega-tool and token passthrough; records wrong annotations, missing audience and tenant enforcement, unbounded disclosure, destructive confirmation, compatibility, and negative-test gaps; links every critical issue and gate; returns a blocked verdict with exact blocker IDs; and names one non-mutating next evidence-producing action that passes the bundled checker.
Creator
Nnalrio7
