Overmind

Open-source (AGPL-3.0) platform that turns your agents' production traces into specialised models you own. Send OpenTelemetry traces or use the Python SDK; your coding agent maps the agent's prompts, tools and control flow, Overmind generates evals from that map, the Data Workshop turns traces into versioned eval and training sets, and optimisers return prompt, tool or code changes as git diffs. Fine-tune an open-weight model (LoRA or full), benchmark it against your current model on the same evals, then serve it behind an OpenAI-compatible API or download the weights. Works from Cursor, Claude Code, Codex or OpenCode via MCP. Free plan (2 training and 2 deploy jobs/mo, $50 one-time credit); Pro $120/mo; training and inference use credits.