Inspiration
Trialrun.dev was sparked by the frustration of context-switching between coding, research, and AI tooling. We wanted a unified workspace where engineers could prototype ideas and instantly wire up MCP tools without DevOps headaches.
What it does
The platform spins up an on-demand sandbox that auto-installs, normalizes, and hot-reloads any MCP server, then exposes them through a FastAPI backend with streaming Claude support. From a single chat box you can query live data, generate images, or run code—all in the same session.
How we built it
We wrapped the official MCP Python SDK with a multi-client manager, added a generic run_mcp.py launcher to handle any repo, and glued it together with FastAPI + AsyncAnthropic for real-time chat. A lightweight React frontend consumes SSE endpoints to display incremental model responses and tool results.
Challenges we ran into
Handling dozens of heterogeneous MCP servers meant solving race conditions, time-outs, and version conflicts, especially with long-running npx processes. Ensuring Claude always received valid JSON-schema tool definitions (and didn’t hallucinate built-ins) took several iterative bug hunts.
Accomplishments that we’re proud of
Boot time for a fresh sandbox—including Git clone and pip install—averages under 15 s, and hot-reloading new servers requires zero restarts. We also built a self-healing config watcher that re-initializes clients when the central handshake changes.
What we learned
Clear, minimal JSON contracts beat clever heuristics when orchestrating LLM tool calls. We also discovered that giving users a simple /configured endpoint empowers rapid experimentation far more than pre-baking every server.
What’s next for Trialrun.dev
We’re adding one-click sharing so a sandbox (with its tool stack) can be published as a public endpoint. In the long term, we’ll ship a marketplace of vetted MCP servers so teams can drag-and-drop capabilities without touching code.
Log in or sign up for Devpost to join the conversation.