Overmind
When an agent fails in production, Overmind scores the trace against your metrics and trains a specialized model on the evidence. Point it at the agent's repository and it maps capabilities, prompts, tools, and tasks into a graph, then matches every incoming OpenTelemetry trace to the capability that produced it and scores it on arrival. Auto-instrumentation picks up the LLM SDKs you already use, any OTel exporter can post to the API, and existing traces from Langfuse, LangSmith, Braintrust, or Galileo sync through connectors. Production traces and uploaded files become versioned datasets, with local English OCR for scanned PDFs and images, and optimisers run prompt, tool, and control-flow experiments in your repository and hand back the winner as a git diff. You fine-tune models whose weights you own, backtest them against the agent's own traces, and serve them, alongside optional OpenRouter frontier models, through one OpenAI-compatible API. Cursor, Claude Code, OpenCode, and Codex drive the whole loop through an MCP server and /overmind commands that cover tracing, datasets, fine-tuning, optimising, and backtesting. Self-hosting keeps traces and training data in your network: one Docker Compose file starts it, a seed command loads a demo workspace, and your own OpenRouter, Modal, and AWS keys power judging, training, and checkpoints. Running on a dedicated VPS on RepoCloud with guaranteed CPU, RAM, and SSD, full root SSH access, and a browser serial console. AGPL-3.0 licensed.
Deploy