Case study
Uplink MCP File Automation
Shipped an MVP-style Claude Code for files: a research CLI for testing MCP workflow limits, filesystem-tool safety, and agent design from no-library prototypes to SDK-based builds.
The work
Uplink is an open-source research CLI for file automation: an MVP-style Claude Code for files. I built it to explore how LLM agents can use tools through the Model Context Protocol, and where those workflows run into practical limits.
How I approached it
I built a TypeScript host application and agents that call MCP tools, including a custom remote storage server with validated tool schemas. The research covered local agents using Ollama and remote agents using the Vercel AI SDK, from no-library prototypes to SDK-based builds.
I evaluated token usage and how increasing the number of tools affected agent workflows. Filesystem-tool safety and the differences between local and remote execution were part of that work.
Results
- Natural-language file operations
- Measured token and tool-scaling limits
- Reusable local and remote workflows
The project is available on GitHub. The screenshot shows its file automation tools in MCP Inspector.