Overmind
A platform that turns an AI agent's own production traces into a fine-tuned model, benchmarked against user-defined evaluation metrics and served behind one OpenAI-compatible API, available hosted or self-hosted.
Overmind trains and improves AI agents using data from their own production traces. Pointed at an agent's codebase, it scans the repository into a graph of the agent's capabilities, prompts, tools and tasks, ingests OpenTelemetry traces and scores them against the capability that produced them, and turns production traces or uploaded files into versioned evaluation and training datasets.
From those datasets it runs prompt, tool and control-flow optimization experiments recorded as a git diff, and produces fine-tuned models that a user owns and can download, retrain or roll back, benchmarked against production traffic and served alongside frontier models on one OpenAI-compatible inference API. It is available from a hosted console, an overmind CLI, a REST API, and an MCP server for Cursor, Claude Code, OpenCode and Codex, driven through slash commands for tracing, dataset management, fine-tuning, optimization and backtesting.
It can run hosted at console.overmindlab.ai or self-hosted with the same codebase, and is written in Python.