overmind

The Training Platform for Specialized Agents

Overmind continuously trains & improves your agents, with data from your production traces

The SDK and CLI you install (pip install overmind) are MIT. The platform behind them is AGPL-3.0 and you can self-host it. Details in Licence.

Point it at your agent's codebase and it turns production traces (or any dataset) into a fine-tuned model, benchmarked against the eval metrics you define and served via 1 unified API, with no ML infrastructure to build.

The weights are yours to download, retrain or roll back.

Available from the Console, the overmind CLI, the REST API, and an MCP server for Cursor, Claude Code, OpenCode and Codex. Hosted at console.overmindlab.ai or run it yourself.

Get started

Hosted

Sign up at console.overmindlab.ai, copy the onboarding prompt from Get started, and paste it into your coding agent (Cursor, Claude Code, OpenCode or Codex) with your agent's repo root open. The Data tab starts from an uploaded dataset instead. It installs overmind, runs overmind init and overmind sync, and maps the agent's capabilities. From then on everything is a /overmind command in the same chat:

Command	What it does
/overmind ensure-tracing	Inspect traces and instrument the agent
/overmind dataset	Build, clean, upload or export a dataset
/overmind finetune	Fine-tune, deploy and smoke-test a model
/overmind optimise	Run prompt and code optimisation
/overmind backtest	Compare models against the agent's own traces

Run it yourself

Self-hosting keeps traces and training data inside your own network. The hosted and self-hosted stacks are the same code.

git clone https://github.com/overmind-core/overmind.git && cd overmind
cp .env.example .env    # set the required keys below
docker compose up -d    # Postgres, Redis, API on :8000, Console on :5173, Celery workers, beat, Workshop sandbox, Grafana on :3001

Document uploads include local English OCR for scanned PDFs, embedded images and direct PNG/JPEG/WebP uploads. Direct images are capped at 100 MB and 64 megapixels; animated images are not supported. The Docker image includes Tesseract and its English/orientation data. For workers or backend tests outside Docker, install tesseract-ocr, tesseract-ocr-eng and tesseract-ocr-osd on Debian/Ubuntu, or brew install tesseract on macOS. No external OCR service is required.

On first boot the API runs migrations; Swagger is at /api/docs/. Sign in at http://localhost:5173 with any email and password. docker compose exec api python manage.py seed_demo --owner <your email> loads a full demo workspace.

The API refuses to start until every required key is set, and the error names each missing one. .env.example documents every key.

Required	Used for
OPENROUTER_API_KEY	Judges, evals, trace scoring, optimiser scoring, frontier inference
MODAL_TOKEN_ID, MODAL_TOKEN_SECRET	Modal training and serving workers (set MODAL_ENVIRONMENT explicitly)
INFERENCE_API_URL, INFERENCE_API_KEY	Serving trained models — the endpoint printed by modal deploy and its shared secret
AWS_ACCESS_KEY_ID, AWS_SECRET_ACCESS_KEY, AWS_BUCKET_NAME	The fine-tuning checkpoint archive
HF_TOKEN	Gated Hugging Face base models; also set it in the Modal secret
SANDBOX_URL, SANDBOX_TOKEN	The isolated Workshop sandbox; .env.example and compose set local values — change the token in production

Optional	Effect
STRIPE_SECRET_KEY	Paid plans and the credit cap; without it, usage is metered with no cap

Fine-tuning and serving also need the Modal workers deployed (modal deploy overbae/modal/modal_vllm_worker.py, register_model.py and modal_sft_worker.py; native decision models also use modal_decision_evaluation.py and modal_decision_serving.py) and a Modal secret named overmind-inference with the AWS keys, INFERENCE_API_KEY and HF_TOKEN.

Self-hosted login uses email and password. Keep Clerk and Stripe secrets blank. Data Workshop transformations are authored by your native coding agent through MCP; the platform stores source files, immutable cell versions and execution receipts. There is no platform Workshop agent or ChatGPT account connection. Reusable scripts, cross-source revisions and continuous bindings use the dedicated Workshop runtime and package protocol. See self-hosting for configuration and private VM setup.

Send a first trace

pip install "overmind[tracing]"
export OVERMIND_API_KEY=ovr_…   # project key from the project page in the Console; add OVERMIND_API_URL for self-host

import overmind

overmind.init(
    service_name="support-agent", capability_id="<capability-uuid>", providers="auto"
)


@overmind.tool()
def search(query: str) -> list[dict]: ...


def handle(request: dict, session_id: str) -> dict:
    with overmind.run(
        "support-run", intent=request["question"], conversation_id=session_id
    ) as run:
        answer = agent(request)
        run.deliver(answer)  # the final output that gets scored
        return answer

providers="auto" instruments the LLM SDKs you already use over OpenTelemetry; without a key, tracing is off and nothing breaks. Any OTel exporter can POST /api/v1/traces instead, and existing traces in Langfuse, LangSmith, Braintrust or Galileo can be synced through a connector. Open Observability to see the trace and its score.

Connect your coding agent

Overmind ships an MCP server at /api/mcp/ with tools, resources and prompts for the platform's agent workflows. Account API keys and OAuth connections can access every active project their account is authorized to use: call list_projects, then pass the selected project_id on project tools and resource URIs. Membership is checked on every operation. Project API keys remain limited to their configured project. Every tool declares what it costs to run (free, compute, llm, gpu) and whether it is destructive. Evaluation matrix rows can be deleted through their scoped lifecycle tool.

The optional plugin packages the MCP connection, Overmind branding and the Overmind workflow skills. OAuth connections remain authorized until revoked; access tokens expire after one hour and refresh tokens rotate. Direct MCP connections can use an API key without OAuth or a plugin.

overmind init --ide <cursor|claude|opencode|codex> prepares the local configuration and overmind sync installs the repository's project API key. To configure an API key by hand:

{
  "mcpServers": {
    "overmind": {
      "url": "https://api.overmindlab.ai/api/mcp/",
      "headers": { "X-Api-Key": "ovr_…" }
    }
  }
}

claude mcp add --transport http overmind https://api.overmindlab.ai/api/mcp/ --header "X-Api-Key: ovr_…"

{
  "$schema": "https://opencode.ai/config.json",
  "mcp": {
    "overmind": {
      "type": "remote",
      "url": "https://api.overmindlab.ai/api/mcp/",
      "enabled": true,
      "headers": { "X-Api-Key": "ovr_…" }
    }
  }
}

[mcp_servers.overmind]
url = "https://api.overmindlab.ai/api/mcp/"
http_headers = { "X-Api-Key" = "ovr_…" }

Domain	Main tools
Projects	list_projects, create_api_key
Observability	inspect_capability_health, query_failures, query_traces, list_signals, set_signal_subscription, get_job, inspect_operation
Datasets	list_datasets, inspect_dataset, query_dataset, create_dataset_from_traces, create_dataset_from_llm_calls, explore_dataset, create_data_partition, inspect_dataset_workbench, save_dataset_pipeline, run_dataset_pipeline, import_dataset_version, update_dataset
Evaluations	check_evaluation_readiness, create_eval_set, create_eval_matrix_row, delete_eval_matrix_row, run_evaluation, compare_evaluations
Training	check_finetune_readiness, prepare_training_data, estimate_finetune, start_finetune, inspect_training_progress, cancel_finetune, deploy_finetune, retry_deployment, set_active_model, run_inference, get_model_swap_prompt
Decision models	list_decision_models, start_decision_job, create_native_evaluation, inspect_native_evaluation, create_training_experiment, measure_decision_performance
Optimiser	check_optimizer_readiness, start_optimizer, inspect_optimizer_result
Connectors	inspect_connectors, configure_connector, sync_connector
Instrumentation	get_instrumentation_plan, verify_instrumentation
Catalog	get_model_catalog, list_model_workflows

tools/list returns the tools the connection's permissions allow: a read-only connection omits every write tool.

Prompts such as investigate-capability, finetune-capability and ship-model chain the tools into complete workflows.

For a self-hosted instance, replace the host with your API URL (http://localhost:8000 locally). Keys are written to git-ignored files only; overmind sync will not write a key into a tracked file.

---

Contributing

Open an issue, or a PR from a feature branch using .github/PULL_REQUEST_TEMPLATE.md — main is protected and AGENTS.md describes how we work. Questions go to the Discord.

---

Telemetry

The SDK and CLI send anonymous usage analytics to PostHog — one cli.invoked event per CLI run and sdk_init on library use; never prompts, trace contents, keys or dataset contents. Opt out with OVERMIND_ANALYTICS_ENABLED=false or DO_NOT_TRACK=1; analytics is also off when CI is set. Your traces go only to your own project.

---

Licence

This repository contains two licences. The platform (overbae/, frontend/, and everything else outside overmind/) is AGPL-3.0; the root LICENSE file is the verbatim AGPL-3.0 text. The SDK, CLI and client libraries under overmind/ are MIT. Copyright (c) 2026 Overmind Ltd. A commercial licence is a paid alternative from Overmind Ltd if you need different terms — [email protected].