Overmind

Overview

Overmind is a platform for continuously improving AI agents. It connects to an agent's codebase and production traces, builds versioned datasets, evaluates behavior, fine-tunes open models, and serves trained and frontier models through one OpenAI-compatible API. Teams can use the hosted console or self-host the platform.

Key Features

• Agent and capability graph scanned from the repository, covering prompts, tools, and tasks.
• Observability with OpenTelemetry traces scored as they arrive and matched to the capability that produced them.
• Datasets built from production traces or uploaded files as versioned eval and training datasets.
• Evaluation and optimisers to measure quality per capability and test prompt, tool, and control-flow changes.
• Model fine-tuning on your own data with weights you own, plus inference for trained and frontier models on one OpenAI-compatible API.
• MCP server and coding-agent workflows for Cursor, Claude Code, OpenCode, and Codex.

Use Cases

• Improve agents handling customer support or coding tasks by using real production traces rather than relying only on prompt intuition.
• Find the cause of agent failures by inspecting the failing step, tool call, and context behind it.
• Build training and evaluation datasets from production traces to capture knowledge specific to a business.
• Prove prompt or model improvements against real tasks and catch regressions before deployment.

Getting Started

• Hosted: sign up at console.overmindlab.ai, pick a coding agent on Get started, and paste the onboarding prompt with the agent repository open. Onboarding installs overmind, runs overmind init and overmind sync, and builds the context graph.
• Use /overmind commands such as ensure-tracing, dataset, finetune, optimise, and backtest in the coding-agent chat.
• Self-host: clone the repository, copy .env.example to .env, set the required keys, and run docker compose up -d.

Deployment & Requirements

• Hosted console is available at console.overmindlab.ai; self-hosting runs the same stack with Docker Compose, including Postgres, Redis, API, console, Celery workers, beat, and Grafana.
• The API requires configured keys for OpenRouter, Modal, an inference endpoint, an AWS bucket, and Hugging Face before it starts.
• Fine-tuning and serving require deployed Modal workers and a Modal secret named overmind-inference with AWS keys, inference key, and Hugging Face token.
• Local document OCR uses Tesseract with English and orientation data in the Docker image; no external OCR service is required.

Before You Adopt

• License: AGPL-3.0. Review its terms before using, modifying, or distributing the project.
• Self-hosting requires all required API keys to be configured before the API will start.
• Fine-tuning and serving depend on deployed Modal workers.
• The SDK and CLI send anonymous usage analytics to PostHog, so teams should evaluate this telemetry behavior.
• Self-hosted authentication uses email and password, and Clerk and Stripe secrets should remain blank.