Detailed Analysis
A freelance software consultant and longtime Claude user describes building KitStack, a platform-as-a-service designed to let developers embed structured, persistent applications directly inside Claude's chat interface, bypassing what the author characterizes as the significant developer-experience shortcomings of the raw Model Context Protocol. The project originated from a concrete operational pain point: professional data — client contacts, meeting notes, project progress, and expenses — was fragmented across Apple Notes, Google Sheets, Gmail, and ephemeral Claude sessions, with Claude's native memory feature proving unreliable for structured retrieval. After observing that some MCP applications were already capable of rendering interactive views within Claude's chat UI, the author recognized an architectural opportunity to build a lightweight CRM that lived natively inside the AI assistant rather than requiring a separate subscription tool.
The core technical contribution of KitStack is a deliberate compression of the MCP surface area. Standard MCP implementations are verbose, token-intensive, and slow — problems the author attributes to the protocol's fundamental design as a heavily structured tool-call mechanism. KitStack collapses the entire MCP interface down to two tools, `kit` and `kit_view`, instructing Claude to treat `kit` analogously to a bash shell. Individual applications, called "kits," employ progressive disclosure so the language model receives only the context it requires at any given moment. A reverse proxy architecture allows a single server to route requests to arbitrarily many kits, meaning developers avoid the overhead of spinning up a dedicated MCP server per application. The platform handles MCP protocol compliance, deployment, hosting, authentication, and database persistence, leaving developers to focus exclusively on business logic and UI.
The broader significance of this project lies in what it reveals about a growing developer pattern: using Claude not merely as a conversational assistant but as a persistent, stateful operating environment. The author's frustration with Claude's markdown-file memory system — which is pulled into context inconsistently and without user control — reflects a widely noted limitation of the current Claude ecosystem. Rather than waiting for Anthropic to solve structured persistence natively, a class of technically capable users is building bespoke infrastructure on top of MCP to achieve it independently. KitStack is an early example of what might be called the "Claude-native application" category, where the AI interface itself is the deployment target rather than a web browser or mobile device.
This development connects to a larger trend of developer tooling being built around Claude's extensibility features, particularly MCP, which Anthropic introduced as a standardized protocol for connecting language models to external tools and data sources. The proliferation of MCP servers and now MCP application platforms signals that the ecosystem is maturing past simple chatbot use cases toward something closer to a programmable operating environment. KitStack's design philosophy — minimizing token overhead, treating the LLM as a command interpreter, and abstracting away infrastructure — anticipates an architectural direction that will likely become increasingly common as developers seek to build durable, low-cost business tooling on top of foundation models without committing to traditional SaaS economics.
Read original article →