AI Prompts for MCP Server Setup

I want to build an MCP server for [YOUR_PROJECT] that gives my Claude Code agent access to [DESCRIBE WHAT YOU WANT THE AGENT TO ACCESS, e.g. GitHub issues, a Postgres database, our Linear board]. For each type of access, tell me: the tool name I should expose, the input parameters it needs, the output format it should return, and whether it needs read-only or read-write access. Then write a one-paragraph MCP server design spec I can use to start implementation.

Design an MCP server for [YOUR_PROJECT] that gives the agent read and write access to GitHub. The server should expose these tools: list_open_issues (returns issues with labels and assignees), get_issue (returns full issue content by number), create_pr (opens a PR with title, body, and branch), post_pr_comment (posts a comment on a specific PR), and add_label (adds a label to an issue or PR). For each tool, write the JSON schema for inputs and outputs.

Design an MCP server for [YOUR_PROJECT] that gives the agent safe read-only access to the [DATABASE TYPE, e.g. Postgres / Supabase / SQLite] database. The server should: expose a query_table tool that accepts a table name and optional filter, enforce a whitelist of allowed tables the agent may read, never expose credentials in tool responses, and limit result sets to 100 rows by default. Write the server design including the tool schema, the safety constraints, and the connection approach.

Write the design for a single Slack notification tool to add to my MCP server for [YOUR_PROJECT]. The tool should: accept a message string and an optional channel (defaulting to #agent-updates), format the message with a header showing the loop name, current task, and timestamp, and post it to Slack via webhook. Write the tool schema, the webhook request format, and the error handling for failed posts.

Write a README for the MCP server I am building for [YOUR_PROJECT]. The README should cover: what the server does and why it exists, the list of tools it exposes with one-line descriptions, how to configure it in Claude Code's settings, the environment variables it needs, how to run it locally for testing, and any limitations on what the agent is allowed to do with it. Keep it under 300 words.

Write a complete MCP tool definition for [TOOL NAME] that does the following: [DESCRIBE WHAT THE TOOL DOES]. The definition should include: the tool name in snake_case, a one-sentence description the agent will use to decide when to call it, the JSON schema for all input parameters with types and descriptions, the JSON schema for the output, and the error response format. Write this as a TypeScript MCP tool using the @modelcontextprotocol/sdk.

Write the implementation for an MCP tool called get_recent_records for [YOUR_PROJECT] that queries the [TABLE_NAME] table in [DATABASE TYPE]. The tool should: accept a limit parameter (default 10, max 50), a filter parameter (optional key-value pairs), and a sort parameter (field name and direction). Return records as a JSON array. Include: the TypeScript implementation using the MCP SDK, the SQL query construction (with parameterized inputs to prevent injection), and the error handling for connection failures.

Write the TypeScript implementation for an MCP tool called list_github_issues for [YOUR_PROJECT]. The tool should: accept optional parameters for state (open/closed/all), label filter, and limit (default 20); call the GitHub REST API using the repo token from environment variables; return an array of objects with id, number, title, labels, and created_at; and handle rate limiting by returning a clear error message instead of throwing. Write the full implementation.

I want to add a create_linear_ticket tool to my MCP server for [YOUR_PROJECT]. Since this tool creates real data, I need it to be safe to call from an agent loop. Write the implementation with these safety features: a dry_run parameter that previews what would be created without posting it, a confirmation field in the output that the agent must explicitly acknowledge before the ticket is created, and a rate limit of 10 tickets per hour. Write the TypeScript implementation.

I have an existing MCP tool for [YOUR_PROJECT] that sometimes fails silently when the external API is unavailable. Rewrite the tool to handle errors explicitly. The new version should: catch network failures and return a structured error response with code, message, and retryable flag; log failed calls to a local file without blocking the response; and return partial results if only some records could be fetched. Here is the current implementation: [PASTE TOOL CODE].

Write the Claude Code configuration to add my local MCP server for [YOUR_PROJECT]. The server runs on localhost and its entry point is [ENTRY FILE PATH]. Write: the .claude/settings.json entry with the correct mcpServers configuration, the environment variable setup for the server's credentials, the startup command, and the verify-it-is-working test I can run from a Claude Code session to confirm the tools are available.

Write an integration test suite for the MCP server in [YOUR_PROJECT]. The tests should: start the server programmatically, call each tool with valid inputs and assert the response shape, call each tool with invalid inputs and assert the error response, simulate an external API failure and assert the error handling, and clean up any side effects after each test. Write the test file using [TEST FRAMEWORK, e.g. Vitest / Jest].

My MCP tool [TOOL_NAME] is failing when called from my Claude Code agent loop. Here is the error the agent receives: [PASTE ERROR]. Help me debug this. Walk through: what the error message indicates about where the failure occurs, what to check in the server logs, what to verify in the tool input schema, and how to write a minimal reproduction case I can run outside the loop to isolate the issue.

Write the test prompts I should run in a Claude Code session to verify my MCP server for [YOUR_PROJECT] is working correctly before I use it in a loop. For each tool I have defined, write a one-sentence test prompt that calls the tool with realistic inputs and tells me what a successful response looks like. Also write the prompt I should use if I want Claude Code to list all available MCP tools and show me their schemas.

I want my Claude Code loop for [YOUR_PROJECT] to check MCP server health before starting each run. Write: the health check prompt the loop should run first, the expected response format, what the loop should do if the health check fails (skip MCP-dependent tasks, alert me, or abort the run), and the log entry format for MCP call success and failure rates over time.