Overmind CLI
Overmind CLI and MCP server: Claude Code, Codex or Cursor fine-tune a model on your agent's traces, benchmark it on your metrics and serve it behind an API
- 757 ★ on the repository
- Last commit
- AGPL-3.0
Broad access
Why this level
- Sends LLM and tool call traces, and on sync repository data, to the Overmind platform (to their servers in the hosted version)
- Starts paid fine-tuning, evaluations and model deployment, and some MCP tools are marked destructive
- The project key lives in `.overmind/credentials.toml`, and an account key opens every project its owner can reach
Install
- Works in
In your terminal, with SkillFoxx CLI
npx skillfoxx add cli/overmindDetects the agents on your machine, shows the access level and pins the version.
Other ways to install
Access and checks
How we reviewBroad accessThe tool can change data, run commands or handle secrets.
- Sends LLM and tool call traces, and on sync repository data, to the Overmind platform (to their servers in the hosted version)
- Starts paid fine-tuning, evaluations and model deployment, and some MCP tools are marked destructive
- The project key lives in `.overmind/credentials.toml`, and an account key opens every project its owner can reach
This is third-party code. Review the repository files before installing.
What it does
Overmind turns your agent's traces or an uploaded dataset into a fine-tuned model. The `overmind` package installs a CLI and SDK, and `overmind init --ide claude` (also `codex`, `cursor`, `opencode`) connects a coding agent: it places a skill in the repository, adds the Overmind MCP server and prints a sign-in step. The next command, `overmind sync`, creates a project and saves its key to `.overmind/credentials.toml`. From there work runs through `/overmind` commands in the agent chat. `/overmind setup` scans the repository and builds a graph of the agent's capabilities, `ensure-tracing` adds tracing, `dataset` builds datasets, `finetune` trains and deploys a model, `optimise` tries prompts and code, and `backtest` compares models on your traces. The best optimiser result arrives as a regular git diff. The MCP server at `https://api.overmindlab.ai/api/mcp/` gives the agent tools for projects, traces, datasets, evaluations, training and connectors. Each tool declares its run cost (`free`, `compute`, `llm`, `gpu`) and whether it is destructive. Trained models are available through an OpenAI-compatible API and `overmind.Client`. The platform can be self-hosted with `docker compose`.
Who it is for. For teams that already build LLM agents and want a model fine-tuned for their tasks without building training infrastructure, working from Claude Code, Codex, Cursor or OpenCode.
Good fit when
- The agent has production traces and you need to turn them into training and test sets
- You want to compare your model with several frontier ones on real traces with `/overmind backtest`
- You need to fine-tune a model and get an OpenAI-compatible endpoint for it
- You want to optimise prompts and tool calls through experiments that arrive as an ordinary git diff
- Traces and training data must stay inside your network: the platform installs with `docker compose`
Not a fit when
- You have no agent to train: Overmind starts from the traces and goals of a specific application
- You want a quick local check with no sign-up and no third-party clouds: both the cloud and the self-hosted version depend on external services
- You look for a tracing library alone: the SDK with `overmind[tracing]` is only part of the set
- The platform's AGPL-3.0 license does not fit your product: the SDK and CLI are MIT, but the platform is AGPL
Example request
Read https://console.overmindlab.ai/setup.md, connect Overmind to this repository and run /overmind setupLimitations
- The cloud version needs an account at console.overmindlab.ai, with a one-time OAuth sign-in.
- Self-hosting will not start the API without `OPENROUTER_API_KEY`, `MODAL_TOKEN_ID`, `MODAL_TOKEN_SECRET`, AWS storage access and a Hugging Face token. Fine-tuning and serving need deployed Modal workers.
- Paid plans and the credit cap turn on with `STRIPE_SECRET_KEY`. Without it usage is only metered.
- Python 3.10 or newer is required. On Windows install the package into a virtual environment.
- Model training and deployment are paid, and some MCP tools are marked destructive.
How to disable. Remove the `overmind` block from your agent's MCP configuration and the skill from the repository folder, then run `pip uninstall overmind`. Delete the `.overmind/credentials.toml` file with the project key by hand and revoke the key in the Overmind console.
FAQ
How do I connect Overmind to Claude Code?
Run `pip install overmind`, then `overmind init --ide claude` and `overmind sync`. The first command installs a skill and the MCP server, and the second creates a project and saves its key. After the sign-in step that `init` prints, the `/overmind setup` command is available in chat.
What does the Overmind MCP server do?
It gives the agent tools for projects, traces, datasets, evaluations, fine-tuning, the optimiser and connectors. Each tool declares its run cost (`free`, `compute`, `llm`, `gpu`) and a destructive flag. The `tools/list` result depends on the connection's permissions, and a read-only connection has no write tools.
Can I self-host Overmind?
Yes, the platform installs with `docker compose up -d` after `cp .env.example .env`. The API refuses to start until required keys are set: `OPENROUTER_API_KEY`, `MODAL_TOKEN_ID` and `MODAL_TOKEN_SECRET`, `INFERENCE_API_URL` and `INFERENCE_API_KEY`, AWS access and `HF_TOKEN`.
What license does Overmind use?
Two. The platform in the repository root is AGPL-3.0, while the SDK and CLI in the `overmind/` folder are MIT. A paid commercial license with other terms is available from Overmind Ltd.
How do I add tracing to my application?
Install `pip install "overmind[tracing]"`, set `OVERMIND_API_KEY` and call `overmind.init(...)` at process start. Then mark functions with the `@overmind.entry_point()`, `@overmind.tool()` and `@overmind.observe()` decorators. Without a key the calls do nothing.
Discussion
No questions or reviews yet.
Related
CLI toolsA CLI and MCP integration keeping agent decisions and checkpoints in local SQLite so work survives a session change.
A local MCP server that stores skills as versions and lets improved variants go live only after a human review
A plugin that gives an agent the IntelliNode skill and MCP server: ask any model, search, review code, tests, SQL, images and speech
Your own server gives every agent shared memory, AST-level code search and pre-commit checks through one MCP endpoint.
Shared company context for agents over MCP: an event graph with permissions, approvals and audit
One runtime for Claude Code, Codex, OpenCode and more: a run counts as done only when checks the agent did not write pass
Comparison with similar tools
Cards that share the main search query, and how they differ.
| Tool | Official | Repository stars | Main difference |
|---|---|---|---|
| Overmind (this card) | yes | 757 | |
| Promptfoo | no | 27,080 | tests prompts and models with evals, while Overmind also fine-tunes a model on your traces |
| LangSmith skills | no | 166 | skills for LangSmith tracing and evals, with no model training |
README in short
The README describes a training platform for specialized agents: capability graph, observability, datasets, evaluations, optimisers, models and inference. Two ways to run follow, hosted and self-hosted, then coding agent setup with configuration examples for Cursor, Claude Code, OpenCode and Codex, an MCP tool table and a first-trace example. Licenses are split between the platform and the SDK.