Riza

Executes arbitrary code in isolated sandboxed environments supporting Python, JavaScript, Ruby, PHP, Go, Rust, and other languages through Riza's managed runtime infrastructure. The MCP server acts as a bridge, translating code execution requests from LLMs into Riza API calls that handle compilation, execution, and output capture in secure containers with resource limits and timeout enforcement.

Implements the Model Context Protocol (MCP) server specification, allowing Claude and other MCP-compatible LLMs to discover and invoke code execution as a tool through standardized JSON-RPC messaging. The server exposes tools with JSON schemas describing parameters, handles tool call requests from the LLM, executes them via Riza's API, and returns structured results back to the LLM for agentic reasoning.

Provides fine-grained control over code execution context through environment variables, stdin piping, and output capture. The execution engine accepts environment variable dictionaries, stdin input streams, and captures both stdout and stderr separately, enabling complex workflows like piping data between code runs, setting API keys for executed code, and debugging output streams independently.

Enforces execution time limits and resource constraints on all code runs, automatically terminating processes that exceed configured thresholds. The runtime monitors CPU, memory, and wall-clock time, killing runaway processes and returning timeout/resource-exceeded errors to the caller, preventing infinite loops or resource exhaustion attacks from impacting the execution service.

Abstracts away language-specific compilation and runtime setup by automatically detecting the target language, invoking appropriate compilers/interpreters, and handling language-specific quirks. For compiled languages (Go, Rust), the system compiles code before execution; for interpreted languages (Python, JavaScript), it directly executes. The MCP server exposes a unified interface where callers specify language and code, and the runtime handles all setup transparently.

Captures and reports detailed execution failures including compilation errors, runtime exceptions, segmentation faults, and timeout conditions with structured error metadata. The system distinguishes between different failure modes (syntax error, runtime error, timeout, resource limit exceeded) and returns them as structured responses, enabling LLMs and agents to understand why code failed and potentially retry or fix it.

Each code execution runs in a completely isolated, ephemeral environment with no persistent state between runs. The filesystem is temporary and discarded after execution completes, preventing code from one execution from affecting subsequent executions and ensuring complete isolation between different LLM requests or agent steps. This design eliminates state management complexity while guaranteeing security isolation.

Manages Riza API credentials and MCP server configuration through environment variables or configuration files, handling authentication to Riza's API and exposing code execution tools to MCP clients. The server reads configuration at startup, validates credentials, and maintains authenticated connections to Riza's endpoints, abstracting credential management from the MCP client.