Turn OpenAPI Spec into MCP Server in Code-mode.
Simplified.

A standalone binary that translates REST-endpoints into agent-ready TS/JS SDK available via MCP to search SDK-docs and execute Code in a secured VM sandbox.

Get started
$ mcpbus --openapi-spec=./CRM-service-OpenAPI-spec.yaml
➔ Parsing OpenAPI, paths found paths_num=1
➔ 🤖 translating OpenAPI endpoint into TS/JS SDK method
    └─ endpoint method=GET path="/tickets"
    └─ method output JSON schema created
    └─ method input JSON schema created
    └─ adding SDK method method_name=crm.get_tickets
    └─ ✅ method created method_name=crm.get_tickets
➔ JS/TS runtime SDK created jit_enabled=true methods_num=1
➔ Starting MCPBus server mcp_version=2025-11-25
➔ MCP-endpoint url="http://localhost:8080/mcp"
➔ 🚀 MCPBus server is running!

Get Started

Code Mode

MCPBus provides a paradigm shift in how your agents use LLMs and call MCP-tools to access external systems and APIs.

MCPBus runs MCP-server in Code Mode - it loads and parses your OpenAPI specs into TS/JS SDK on the fly and then makes this SDK available to your agents for search and code execution over two MCP-tools (806 tokens context window footprint).

Instead of dealing with dozens (or hundreds) of MCP-tool definitions and guessing what tool to use and how, your agent utilizes LLM for what it can do the best - writing Code.

TS/JS VM

MCPBus creates memory-pinned and governed VM sandbox to run TS/JS code with calls to SDK methods.

Our VM is pure in-process runtime for TS/JS. It supports modern TypeScript or JavaScript, it is highly optimized, protected, and allows to execute only generic TS/JS code with calls to the available SDK methods.

LLMs are strictly grounded by our VM: no file system nor OS-level functions access available, no network access, no Web-API access, no modules. All interaction with outside world can happen only via calls to generated SDK methods.

How it works

  1. MCPBus provides MCP-tool named search_docs to let your agents search SDK documentation and find methods most relevant to user's task.

  2. Your agents use LLM to initiate SDK docs discovery - this might be one call or several calls to search_docs as LLM creates a multistep plan how to solve user's task and reasons what SDK methods are needed.

  3. Then, LLM generates TS/JS code to implement multistep logic, which solves user's task.

  4. Agent uses MCP-tool execute to submit this code to MCPBUs for execution in VM.

  5. Your agent receives result from execute MCP-tool. This result usually represents high-level data products, calculated, filtered and aggregated data.

  6. LLM is provided with result to prepare human friendly reply to the user or to keep it as structured JSON in case it is needed for next step in agentic flow.

Code Mode benefits

  • Code Mode allows LLM to generate solution for the problem in one run, without multiple round trips "LLM -> guessed tool-1 call -> tool-1 result -> LLM -> guessed tool-2 call -> tool-2 result -> LLM -> ... and so on". This approach causes accuracy decay on the individual agent level and just doesn't scale.
  • For most tasks you will need only one universal agent in Code Mode, which means no accuracy decay causing agentic drift
  • Code Mode agentic AI flow accuracy is above 95%
  • Token consumption is a fraction of what you used to have, Code Mode slashes token expenses by up to 98%
  • No corp data leakage to LLM providers. With MCPBus, you never submit raw output to LLM to decide what to do next.

Live example

If you are interested to see Code Mode in action - please have a look at MCPBus NWS showcase, a publicly available MCP-server ran with MCPBus, which provides access to National Weather Service API.

Next
1. Download & Install