What it does
NitroStack is a developer platform for building and running Model Context Protocol (MCP) servers and MCP apps. It has four parts: an open-source TypeScript SDK and CLI for writing MCP tools, resources, prompts and UI widgets with decorators and Zod schemas; NitroStudio, a desktop app for inspecting, testing and chatting with MCP servers; NitroCloud, serverless hosting that deploys MCP servers from GitHub; and NitroChat, a white-label AI chat client for MCP servers. It is operated by Nitrostack Inc., a Delaware corporation.
Official source 1 ↗ · Official source 2 ↗Features
- Decorator-based TypeScript SDK. MCP tools, resources and prompts are declared with decorators and Zod input schemas, organised in modules with dependency injection, in a structure modelled on NestJS. Official source 1 ↗ · Official source 2 ↗
- Built-in auth, caching and rate limits. Tools can be guarded with API key, JWT or OAuth 2.1 authentication and role- or attribute-based access control, and decorated with caching, rate limiting and middleware. Official source 1 ↗
- UI widgets for tool output. A tool can be linked to a React/Next.js widget that renders its output as an interactive interface in compatible clients instead of plain text. Official source 1 ↗ · Official source 2 ↗
- CLI with templates and code generation. The CLI scaffolds projects from five official templates, runs a hot-reload dev server, generates modules, tools, guards and types, and builds production bundles. Official source 1 ↗
- Generate MCP servers from REST APIs. NitroStudio can turn a Swagger/OpenAPI URL or a Postman Collection into a scaffolded NitroStack MCP project, using AI to detect endpoints and generate typed tools. Official source 1 ↗
- MCP inspector and AI chat testing. NitroStudio lists every tool, resource, prompt and task a server exposes, runs them from forms, previews widgets, and offers an AI chat that calls the server's tools in real time. Official source 1 ↗
- Ops Canvas and AI-generated test cases. Ops Canvas draws a graph of which tools the AI called, in what order and how long each took, with per-turn token usage; test cases define expected tool calls and can be generated from tool schemas. Official source 1 ↗
- Marketplace and multi-server sessions. NitroStudio connects to pre-built MCP servers such as GitHub, Figma, Linear, Jira, Supabase and Vercel, and can open several servers at once in tabs. Official source 1 ↗
- Voice mode. NitroStudio has speech-to-text and ElevenLabs text-to-speech in seven languages for hands-free testing. Official source 1 ↗
- NitroCloud serverless hosting. NitroCloud deploys MCP servers and MCP apps from GitHub with automatic scaling, SSL and, on Pro, custom domains, log retention and usage analytics. Official source 1 ↗ · Official source 2 ↗
- NitroChat white-label client. NitroChat is a brandable AI chat front end for MCP servers that supports several AI providers. Official source 1 ↗
Pricing & free limits
30k requests a month at 30 requests a minute (burst 10 req/s for 30s), scale-to-zero compute with cold starts, a hard cut at the limit, GitHub deploy, NitroStudio, usage dashboard and community support. No credit card required. Official source 1 ↗
300k requests a month at 300 requests a minute (burst 50 req/s for 60s), no cold starts with one always-on instance and autoscaling up to 3 instances, overage at $0.10 per 1,000 requests up to 10M extra, custom domains, 7-day log retention, usage analytics and email support. Official source 1 ↗
Dedicated nodes, custom request quotas and rate limits, a 99.9% uptime SLA, volume overage discounts, RBAC, audit logs, custom log retention, dedicated Slack support and onboarding. Official source 1 ↗
Billing details
Pro overage is charged only when usage-based billing is enabled. A request is any HTTP call to the MCP server, counted at ingress, so retries count separately; HTTP 429 responses are not counted. Official source 1 ↗
AI model use is billed per million tokens at listed rates for Anthropic Claude, Google Gemini, OpenAI GPT and xAI Grok models; for example Claude Sonnet is listed at $2.00 input and $10.00 output. Official source 1 ↗
Free access & limits
Hobby allows 30k requests a month with a hard cut at the limit (no overage), 30 requests a minute, and scale-to-zero instances with cold starts of about 3–6 seconds. Official source 1 ↗
Best for & limitations
Wrapping an existing REST API as an MCP server so AI clients can call it, by importing its OpenAPI spec or Postman Collection into NitroStudio. Official source 1 ↗
Building an infrastructure MCP server with read-only diagnostic tools and one gated action: the EdgeSentinel example diagnoses Linux edge devices, was tested in NitroStudio's AI chat and is served from a NitroCloud endpoint. Official source 1 ↗
Adding a conversational front end to an existing agent backend: WeKan used NitroStack as the MCP runtime that lets a chat ask questions, save answers and trigger its existing AI workflows. Official source 1 ↗
- TypeScript only for now. The SDK is a TypeScript framework; the docs list Go and Rust as coming soon. Official source 1 ↗ · Official source 2 ↗
- Free tier cold starts and hard cap. Hobby servers scale to zero with cold starts of about 3–6 seconds and stop serving at 30k requests a month; always-on instances and overage need Pro. Official source 1 ↗
- AI requests go to third-party providers. AI prompts and inputs are routed through OpenRouter to the selected model provider, whose retention and training practices may differ; NitroStack does not use customer data to train generalised models. Official source 1 ↗
- Short log retention. Pro keeps 7 days of logs, and NitroStack's standard operational log retention is 14 days. Official source 1 ↗ · Official source 2 ↗
Submit a tool
