React Toolkit MCP server

mcp-toolkit

MCP servers for React + TypeScript development automation. Works with Claude Desktop, Cline, Cursor — and as plain CLI scripts — one protocol, zero duplication.

59 tools across 9 categories — component scaffolding, code quality, a full CRUD-feature factory, CRA→Vite migration, and more. Every tool ships as its own MCP server, built and tested independently.

Why this exists — the token math

Here's the thing nobody tells you when you start building agentic workflows: the loop itself is what's expensive, not the model. An agent working without any composed tools does everything the slow way — read a file, think, write a file, read it back to check its own work, repeat — and every single one of those turns re-sends the whole conversation so far as input tokens. By the time you're 20 steps into a real multi-file task, that resent context alone can be running 50K+ tokens per call. It adds up fast, and it's not really about which model you're using.

I didn't just take that on faith — a few sources back it up with real numbers. LeanOps measured agent loops running about 3.2× the tokens of a single direct call at 5 steps, ~30× at 50 steps, and past 100× once you're deep into a typical build-and-debug session — because re-sent context is roughly 62% of the bill. Vantage found similar: real agentic sessions run an input-to-output ratio around 25:1 (a direct call is closer to 1:1), with a 50-turn session routinely hitting a million input tokens, and non-agentic usage on comparable work costing something like 200× less per interaction on the same team. And a recent arXiv paper on agentic tokenomics puts agentic tasks at roughly 1000× the tokens of single-turn work, with up to 30× variance run to run on the exact same task — so it's not just expensive, it's unpredictable.

That's the problem this toolkit's composed tools are built to get rid of. workflow-runner's schematofeature and cra-to-vite don't add one more tool call into an agent's existing loop — they replace what would otherwise be 7 or 8 separate read/write/verify turns with a single in-process call that runs the whole generator or migration pipeline and hands back the finished result. That's the "50-turn loop collapses into 1 call" shape the research above says saves 10–100×, which is a very different thing from just bolting one extra tool onto an unchanged loop (that only gets you the 20–40% range).

To keep myself honest, I also ran a real, measured benchmark rather than just trusting the theory — ax-benchmark, 6 tasks, claude -p running headless, three arms (agent alone, agent with one MCP tool call added into its loop, and the tool called directly with no agent at all). This is a conservative baseline on purpose, since it only tests adding a single tool call into an otherwise unchanged loop, not the deeper pipeline collapse described above:

Shortened. The full README is on GitHub.