Last updated: 2026-10-13
Overmind is a training platform for AI agents, open-sourced under AGPL-3.0 on 1 October 2026. It captures an agent's production traces over OpenTelemetry, scores them, turns them into versioned datasets and runs LoRA or full fine-tuning of an open-weight model, then serves the result on an OpenAI-compatible endpoint with downloadable weights.
About Overmind
Overmind comes from Overmind Technology Ltd, a London company founded in 2025 by CEO Tyler Edwards, who previously worked in British intelligence at MI5 and MI6, and COO Sam Brunt, an operator who helped scale Funding Circle, Pipe and Vertice. It launched in February 2026 with a £2 million seed round led by Osney Capital, with 14Peaks, Portfolio Ventures, Antler and Endurance Ventures joining. Coverage of that round, in tech.eu, described the product as a supervision layer for AI agents. By autumn it had repositioned as a training platform and published its code.
The idea is that most agents call a frontier model for jobs a much smaller model could learn. Overmind collects what the agent actually does in production, turns those runs into training and evaluation data, and fine-tunes an open-weight model on them. Its catalogue runs from 230M-parameter LFM2.5 up to Qwen 2.5 72B and includes Qwen3.8-27B, Qwen3.6-27B, Gemma 4 12B, Gemma 4 31B and Nemotron 3.5 Lightning; Qwen models from Alibaba Cloud make up more than half of the list. The new model is then measured against the one already serving traffic, so a switch is a decision backed by scores rather than a hunch.
Most of the work is meant to happen inside a coding agent rather than a dashboard. After a one-time onboarding prompt, Claude Code, OpenAI Codex or Cursor can instrument the agent, build a dataset and launch a fine-tune through slash commands. Under the hood the platform leans on OpenRouter for its judges and on Modal GPUs for training and serving, and gated base models come from Hugging Face.
It is aimed at small engineering teams with an agent already in production and one narrow, high-volume task, such as routing, extraction or labelling, where a frontier model is overkill. Teams that already hold clean datasets and only want raw fine-tuning capacity may prefer Together AI or Fireworks AI, and teams after testing and governance without training should compare Openlayer instead. More options sit in the machine learning category, and tracing-first tools are grouped under agent observability. HokAI's guide to the three jobs agent observability tools do explains where tracing alone stops being enough.
Pricing
Free at $0 a month and Pro at $120 a month; Enterprise is quoted per contract and adds SSO, dedicated support and a compliance scope agreed in the contract. Every plan draws on a prepaid credit balance for training, hosted inference, Data Workshop chat and optimizer runs, and new accounts start with a one-time $50 credit. Sending traces never uses credits.
ai/models, read 13 October 2026). No per-credit or per-GPU-hour rate card is published.
| Tier | Monthly price | What it includes |
|---|---|---|
| Free | Free | Fair-use traces, unlimited evals, monthly caps of 10 optimizer runs and 2 each of training and deploy jobs, 5 projects, $50 one-time credit |
| Pro | $120/mo | Unlimited optimize runs, training jobs, deploys and projects, priority support; usage billed from credits |
| Enterprise | Custom | Everything in Pro plus SSO and custom auth, dedicated technical support, custom onboarding and per-contract compliance scope |
Key Features
- OpenTelemetry trace capture: Setting providers="auto" in the Python SDK instruments the LLM client libraries an agent already uses, any OTel exporter can post spans to the same endpoint, and connectors backfill traces from Langfuse, LangSmith, Braintrust or Galileo.
- Capability map from the repo: A local scan of the agent's codebase builds a graph of its capabilities, prompts and tools, and each incoming trace is scored and matched to the capability that produced it.
- Data Workshop datasets: Production traces or uploaded JSONL files become versioned train and eval datasets, and launching a run pins the exact dataset versions it trained and was scored on.
- Before-and-after benchmark: A run can score the untouched base model, the trained checkpoint and the production incumbent on one pinned eval set with frozen graders, adding per-class precision, recall, F1 and a confusion matrix for labelling tasks.
- LoRA or full fine-tuning: Smaller dense models offer a choice of LoRA adapters or full fine-tuning while larger dense and all mixture-of-experts models train as LoRA, and one launch can run several experiments with different learning rates or base models side by side.
- Hosted serving and weight download: A finished job deploys to an OpenAI-compatible chat completions endpoint, can be swapped in behind a capability alias, and its checkpoint downloads from the console or with overmind model download-checkpoint.
- MCP server for coding agents: 36 MCP tools cover tracing, datasets, evals, training and the optimiser, each tool declares its cost class, and none of them can delete anything.
- Self-hostable platform: The platform is AGPL-3.0 and boots with docker compose (Postgres, Redis, API, console, Celery workers and Grafana), while the SDK and CLI ship under MIT.
Pros
- Ownership goes beyond a slogan: the code can run on your own infrastructure and a trained model can be served somewhere other than Overmind.
- Evaluation is built to resist flattering numbers: overlap between train and eval data is flagged, token-limited outputs are marked incomplete instead of scored, and a run on degraded evidence reports failure with no headline delta.
- Low switching cost for teams already tracing elsewhere, because existing observability data can be imported instead of re-instrumenting the agent first.
- Budgeting a first run is straightforward: the setup wizard estimates cost and duration per base model before launch, and every experiment reports what it actually cost when it ends.
Cons
- Supervised fine-tuning only, with no reinforcement learning over multi-step trajectories; Overmind's OpenPipe comparison page sends RL users to the ART trainer or W&B Training.
- A self-hosted install is not air-gapped by default: the API will not boot without OpenRouter, Modal and AWS keys, so judging, GPU work and checkpoint storage still run on third-party accounts.
- There is no TypeScript tracing SDK, so Node.js agents must be instrumented with plain OpenTelemetry against the OTLP endpoint.
- The plan fee is not the whole bill: usage credits are charged on top, and with no published rate per credit the monthly cost depends on how much you train and serve.
- A young product from a company founded in 2025, with no G2, Trustpilot or Product Hunt reviews yet to check other customers' experience against.
Data Handling
- Training-data policy
- Vendor states customer operational data is not used to train public foundation models. For platform content Overmind acts as data processor under its Terms of Service and Data Processing Addendum.
- Data retention
- 30 days
- Compliance
- SOC 2 (stated by the vendor in its 1 October 2026 press release; the Vanta trust center did not load for verification) · UK GDPR (privacy policy)
Frequently Asked Questions
What does Overmind actually cost?
The Free plan costs nothing and Pro is $120 a month, which removes the monthly caps on optimize, training and deploy jobs and adds priority support; Enterprise is priced per contract. On either self-serve plan, GPU training, model serving, the Workshop chat agent and optimiser jobs also consume prepaid credits, and new paid work is refused once the balance hits zero until you top up. As a sense of scale, the model library lists Qwen 3 8B from $1.00 to train and from $6.00 per million tokens to run.
Does Overmind have a free plan?
Yes. The Free plan covers fair-use trace ingest, unlimited eval runs and, each calendar month, up to 10 optimize runs, 2 training jobs and 2 deploys, across 5 projects with one seat each and community Discord support. A one-time $50 starter credit is what pays for the GPU training and hosted inference on that plan. Running the open-source code yourself is the other free route, though the third-party services it calls bill you directly.
What are Overmind's closest competitors?
The nearest live match is [Maitai](/hub/tools/maitai), which also turns production traffic into customer-specific models, but ties that continuous improvement to its Enterprise plan. Teams that want testing, monitoring and governance with no training at all can try [Openlayer](/hub/tools/openlayer) instead. For reinforcement learning, OpenPipe's open-source ART trainer and W&B Training are the usual picks, because [OpenPipe](/hub/tools/openpipe) itself stopped taking new training and inference on 30 July 2026, according to its own migration notice.
How does Overmind compare to Together AI in 2026?
[Together AI](/hub/tools/together-ai) sells fine-tuning as a prepaid API billed per million training tokens, alongside serverless inference and GPU clusters, and expects you to arrive with a clean dataset and your own evaluation. Overmind starts earlier in the pipeline: it builds the dataset from your agent's traces and scores the fine-tune against the model you already run. Choose Together for raw capacity and transparent per-token rates; choose Overmind when the hard part is knowing what to train on and whether the result is better.
What does it take to start using Overmind?
Create an account at console.overmindlab.ai, pick your coding agent on the Get started page and paste the onboarding prompt into it with your agent's repository open. That prompt installs the overmind package, runs overmind init and overmind sync, and builds the capability graph; after that, /overmind ensure-tracing instruments the agent and /overmind finetune checks readiness, estimates cost and launches the run. A first fine-tune needs a train dataset, a separate eval dataset and at least one enabled grader.
Top Alternatives
- Maitai: Pick Overmind if you want trace-to-fine-tune on a self-serve plan with weights you keep; pick Maitai when a managed control plane should handle model improvement for you on its Enterprise tier.
- OpenPipe: Pick Overmind if you are moving off OpenPipe and want supervised fine-tuning with tracing attached; choose OpenPipe's successor, W&B Training, if reinforcement learning is the requirement.
- Openlayer: Pick Openlayer if risk and compliance teams need AI testing, monitoring and governance with no training involved; pick Overmind when you want your own fine-tune to replace a frontier model.
- Together AI: Pick Together AI if your datasets are already clean and you want per-token fine-tuning capacity at scale; pick Overmind if the traces, datasets and evals still have to be built.
HokAI guides covering Overmind
- Hugging Face vs OpenPipe: Which Should You Use in 2026?: OpenPipe stopped training on 30 July 2026. We compare Hugging Face TRL, CoreWeave Forge, Together AI, Fireworks and Overmind on price, RL support and data.