Skip to main content

Model Context Protocol (MCP)

tip

MCP support requires proxymock be installed. It's free. Instructions are here.

proxymock records API/Database calls going into and out of an app running on your local desktop.

  • inbound requests become tests
  • outbound requests become service mocks (automatically)
  • recordings are portable

You can learn more at proxymock.io

An MCP server is built into proxymock to let Agent-based LLM tools like Cursor, Claude Desktop, VS Code, GitHub Copilot or Amazon Kiro interact. Once installed, you will be able to interact with proxymock using your normal chat interface. This is made possible by the Model Context Protocol.

Installation instructions for proxymock are here (it's really easy)

MCP at a Glance​

MCP is a lightweight, open protocol that standardizes how AI applications (IDEs, chat apps) connect to capability providers (like proxymock). It defines a small set of API types so tools can interoperate:

  • Tools: Callable functions with typed inputs/outputs that models can invoke autonomously (e.g., "start_mock", "record_traffic"). The AI decides when to call these.
  • Prompts: Parameterized prompt templates that users explicitly select, often surfaced as slash commands (e.g., "generate test from capture"). User-controlled, not auto-invoked.
  • Resources: Read‑only documents or data that can be fetched by URI (e.g., local files, logs, recordings) to ground the model with context.
  • Sampling: Allows servers to request LLM completions from the host application, enabling agentic behaviors while maintaining user control.

Together these make agent integrations predictable across different AI applications (Cursor, Claude Desktop, VS Code/Copilot Chat) without custom adapters.

For the full list of tools, prompts, and resources the proxymock MCP server exposes — with parameters and read-only markers — see the MCP Tools & Prompts Reference.

For the agent skills Speedscale publishes and what each one is for, see Agent skills.

Agentic vs IDE Assistants​

  • Agentic tools (e.g., Claude with MCP): Plan multi‑step changes, call tools autonomously, and generate cohesive cross‑file updates. Best for scaffolding features, wiring integrations, or orchestrating proxymock actions end‑to‑end.
  • Interactive IDE tools (e.g., GitHub Copilot, Cursor): Offer fast, in‑editor completions and small, targeted edits with strong local context. Best for iterative refactors, filling functions, and polishing code.
  • Practical workflow: Use agentic generation to draft broader changes or runnable flows, then rely on IDE assistants for tight iteration, review, and final touches.

proxymock provides support for both types of workflows. For fully agentic workflows you should consider using a full sandbox environment with dedicated resources. This will allow your AI agents to run autonomously without your laptop.

Slash Commands vs Natural Language​

When interacting with AI assistants via MCP, you can use both slash commands and natural language to work with proxymock.

  • Slash Commands: These are structured commands that usually start with a / (for example, /record_my_app or /find_breaking_api_changes). In MCP terminology, slash commands typically invoke prompts—predefined templates that you explicitly select. They provide a consistent, repeatable way to trigger specific workflows. Many AI applications (like Cursor or Claude Desktop) surface MCP prompts as slash commands in their UI.

    • Example: 
      /find_breaking_api_changes
      This invokes a guided workflow that helps you compare recorded and replayed traffic to identify API regressions.

  • Natural Language: You can describe what you want in plain language (for example, "run a mock server" or "please record all API calls"). The AI interprets your intent and may automatically invoke MCP tools—capabilities that the model can call autonomously—or combine multiple actions to fulfill your request. This is flexible and works well for exploratory or multi-step tasks.

    • Example: 
      Please start recording all API calls and generate a test for the latest traffic using proxymock

When to use which?

  • Use slash commands when you want a specific, predefined workflow or template applied consistently.
  • Use natural language for exploratory work, multi-step tasks, or when you want the AI to determine the best approach.

Most modern AI applications support both approaches, and MCP enables seamless integration of both interaction styles. Try both to see which fits your workflow best!

🔗 MCP & Protocol Information

Official MCP specification, documentation, and community resources for the Model Context Protocol.
MCP SpecGitHubRegistry

🤖 AI Code Generation & Development

Leading AI-powered development tools that integrate with MCP for enhanced coding productivity.
GitHub CopilotCursorClaudeOpenAI Codex

📚 Learning Resources

Comprehensive guides for MCP integration, AI-assisted development best practices, and proxymock documentation.
MCP GuideBest Practicesproxymock Docs

👥 Community & Support

Join the MCP or Speedscale community, learn about AI developement, and get support for your development needs.
MCP DiscordSlackVibe CodingContext Engineering

Install or remove an agent integration​

Agent setup is optional and separate from proxymock init. It needs no Speedscale account. Select the client and preview the paths before applying changes:

proxymock mcp install --client claude-code --skills --dry-run
proxymock mcp install --client claude-code --skills

Replace claude-code with cursor, claude-desktop, vscode, gemini, opencode, kilo-code, hermes-agent, codex or kiro. The selected client must be detected. Plain proxymock mcp install in a terminal still prompts for each detected client. Interactive confirmation defaults to decline and shows the scope, transport, MCP config path and skills path.

Omit --skills for MCP only. Use --skills-only for bundled skills without MCP. Skills are supported for Claude Code, Codex, Kiro, Kilo Code and Hermes Agent. Other selected clients receive MCP registration only.

Scope and unattended setup​

The default --scope auto prefers the current project for Cursor, Claude Code, VSCode, Gemini, OpenCode, Codex and Kiro. Other clients use user scope, shown before confirmation. Choose --scope project to reject global-only destinations, or --scope user when you deliberately want a global integration. Run project setup from the application repository. Skills support project scope only for Claude Code, Codex and Kiro. Kilo Code and Hermes Agent skills are user scope only, and --scope project is rejected for them.

Project registrations store the bare proxymock command so committed files carry no home-directory paths. Each teammate needs proxymock on their PATH. Decide deliberately whether to commit or ignore .mcp.json and .claude/skills.

For unattended setup, --yes only skips confirmation and never selects clients. In a non-interactive session such as CI or an agent, confirmation is also skipped when --client or --all-clients is given:

proxymock mcp install --client codex --skills --yes
proxymock mcp install --all-clients --yes

The second command explicitly selects every detected client and omits skills. proxymock mcp install --yes without client selection now exits without writes. Migrate scripts to a named client or an explicit --all-clients choice. The older mcp --install flag is deprecated; use the scoped subcommand.

Codex project setup writes .codex/config.toml and optional .agents/skills. The host still decides whether to trust project configuration. Installation never approves hooks or changes host trust settings.

Preserve edits and remove an integration​

Managed installation leaves unowned or edited skills unchanged and refuses to overwrite an unowned or edited proxymock MCP entry. Keep the ownership record next to the client config (the config filename plus .proxymock-install.json). Older installations without a record remain unchanged; use proxymock mcp json for a manual config block instead of deleting unfamiliar entries.

To remove owned artifacts, use the same client and scope as the install:

proxymock mcp uninstall --client claude-code --skills --dry-run
proxymock mcp uninstall --client claude-code --skills

Omit --skills to remove only the MCP registration. Edited files and unrelated client settings remain. Initializing or upgrading the CLI does not silently restore removed integrations.

Network transport and manual setup​

Local stdio is the default. For a remote dev box or container, use matching HTTP settings on the server and installer:

proxymock mcp run --http --port 7799
proxymock mcp install --client claude-code --http --port 7799

The installer writes http://localhost:7799/mcp. If the host runs elsewhere, use proxymock mcp json --http --port 7799 and configure a reachable address in the client. The older --sse transport is deprecated.

Restart the selected agent after setup. For clients the installer does not recognize, proxymock mcp json prints a config block without changing files.