Skip to main content
Overmind ingests OpenTelemetry traces, scores them as they arrive, turns them into curated datasets, and feeds three kinds of work: eval runs, optimisation runs that score code diffs in your repository, and training runs that produce models benchmarked against the one you ship today. The platform is open source. overmind-core/overmind holds the API, the Console, and the SDK, and runs the same way hosted at overmindlab.ai or self-hosted with Docker Compose. Everything speaks standard protocols: OTLP/HTTP for ingest, REST JSON for the API, MCP for coding agents, OpenAI-compatible serving for models. There is no proprietary agent framework. The Python SDK auto-instruments OpenAI, Anthropic, Gemini, Agno, and LangChain over OpenTelemetry, and any OTel exporter can POST to the same endpoint. What the platform does:
  • Live scoring: every trace is carved into task executions and judged against the capability’s eval set
  • Datasets from traffic: traces become eval and training data, shaped by the dataset’s own agent
  • Eval runs: variants measured against a baseline with typed scores
  • Optimiser: candidates generated and replayed in your repository by the CLI, scored on the platform, the winning diff left for you to apply
  • Models: GPU runs on Modal or Baseten, loss curves, a benchmark against your production model
  • Inference: OpenAI-compatible API for your trained models and frontier models
  • MCP and CLI: the same workflows from Cursor, Claude Code, OpenCode, or Codex

​
Surfaces

Authenticate with an ovr_ key as X-Api-Key: <key> or Authorization: Bearer <key>. Keys are scoped to an account or to one project; see Projects & administration.

​
Guides

Quickstart

One pasted prompt: SDK, MCP, capability scan, first trace.

Self-hosting

Docker Compose, the environment reference, backends and billing modes.

Observability

OTLP ingest, span model, attribute mapping, the executions view.

Python SDK

init(), the run boundary, decorators, tasks, runtime expectations.

Agent & Capabilities

Local discovery, the scan push, tasks, telemetry attribution.

Datasets

Source, cells, versions, contracts, the workshop agent.

Eval

Evaluator kinds, claims, task bindings, eval runs.

Trace scoring

Units, tasks, verdicts, execution and session scores.

Optimisers

The client-driven loop, harness and backtesting modes, the winning diff.

MCP

Setup per coding agent, the tool catalog, prompts, resources.

​
Models & API

Models

Runs, the model catalog, loss curves, the baseline benchmark.

Inference

Serving lifecycle, the capability alias, /api/v1/chat/completions.

REST API

Auth, endpoint map, conventions, Swagger.

CLI

overmind init, scan, dataset, optimise, model.