The Problem
For LLM researchers, setting up LLM training or reinforcement learning environment for real-world tool use is complex and painful:- Managing different environment or test accounts
- Implementing MCP Servers and handling various authentication issues
- Initializing realistic data
- Resetting states between multiple runs
- Ensuring isolation across concurrent sessions
The Solution
Klavis MCP Sandbox as a Service solves these challenges. In addition to letting your model interact with our comprehensive MCP server ecosystem, you can use our sandbox infrastructure to easily verify and reset data on any concurrent run.Our sandbox infrastructure is horizontally scalable, so it can handle any number of concurrent sessions as you need.
Lifecycle
Create
Request a sandbox based on the external services you need (Snowflake, GitHub, Notion, Woocommerce, etc.) and get an MCP server URL for that isolated instance.
Initialize
Load a deterministic “world state” in JSON format or via API. We handle everything—creating databases, setting up data, and more.
Interact via MCP
Let your LLM / AI agent use MCP tools against the sandbox as if it were the real app. You can use multiple MCP servers with many tools simultaneously.
Verify
Access the full sandbox state to programmatically compare against your ground truth—whether your LLM completed the task correctly or not.
Video
Resources
Example Notebook
Create sandboxes, seed data, run an agent, then verify and clean up.
Sandbox API
Manage isolated sandbox environments for training/eval: pooling, init, export, teardown.
Fireworks + Klavis
Use Klavis MCP Sandbox with Eval Protocol for model training and RL at scale.