AI Workbench MCP

What AI Workbench MCP MCP server does

AI Workbench MCP MCP server exposes a small, local catalog through the Model Context Protocol. The catalog contains reusable prompt templates and assistant blueprints. It is designed as a narrow, read-only boundary: tools can inspect bundled content or produce rendered results, but they do not modify files, call external providers, execute commands, or access user accounts.

The server offers three tools:

• list_prompts returns the available prompt templates and assistant blueprints.
• render_prompt applies explicitly supplied string variables to a bundled prompt template.
• get_assistant returns one assistant blueprint in a requested supported format.

Assistant output can target ChatGPT, Claude, Gemini, Grok, or the portable Agent Skill format, as supported by the catalog.

How it works

AI Workbench MCP MCP server communicates with an MCP host over standard input and output. The host starts the local executable, performs the MCP handshake, and invokes the exposed tools. No network request is needed for normal operation, and the runtime uses Python’s standard library without additional runtime dependencies.

Each public tool advertises a read-only, non-destructive, idempotent, and closed-world contract. Inputs use bounded, top-level schemas, and unknown tool requests receive bounded errors. These properties make the server suitable for hosts that need a narrowly defined local data source, while still requiring the usual trust review for any MCP server added to an environment.

Setup and configuration

Install the published alpha package with Python 3.10 or newer:

python -m pip install "alptugharun-ai-workbench-mcp==0.1.0a2"

Before configuring a host, run the local diagnostic command:

alptugharun-ai-workbench-mcp --doctor

The diagnostic checks package and Python information, tool names, catalog counts, and the declared read-only boundary. It checks the local package layer only; it does not verify host compatibility.

Configure an MCP client to launch this command as a stdio server:

alptugharun-ai-workbench-mcp

The repository includes a Cursor setup and verification guide. Other hosts may use different configuration formats, so their current MCP documentation should be followed. The package also supports --version for local version checks.

Tools and capabilities

AI Workbench MCP MCP server keeps its capability set intentionally small. list_prompts helps an agent discover what is bundled before selecting an asset. render_prompt accepts explicit string values and returns the resulting prompt without writing the output anywhere. get_assistant retrieves a named blueprint in one of the supported assistant-oriented formats.

The repository provides protocol and tool tests, including a smoke client. Local verification can be run with:

python -m unittest discover -s tests -v
python examples/smoke_client.py

Limitations and notes

The server is read-only and does not provide prompt editing, persistence, file operations, shell access, browser access, provider API calls, or network integrations. Its catalog is local and bundled, so it is not a remote prompt repository or a synchronization service.

Cursor is the host documented and exercised by the project. Successful operation in that host is not a universal compatibility guarantee for every MCP client. The --doctor command confirms local package health, not client configuration or host interoperability.

The project is licensed under MIT. The repository publishes a package to PyPI and identifies its MCP registry entry as io.github.alptugharun/ai-workbench-mcp.