overmind command ships with the Python package. It is the local half of every workflow that touches your filesystem: it installs the skill and MCP entry, syncs overmind.toml, lands files as datasets, holds provider keys for tracing connectors, drives optimisation runs in your repository, and downloads weights. The MCP server never edits a repository or moves bytes; the CLI does.
--api-key, then .overmind/credentials.toml (written by overmind sync), then OVERMIND_API_KEY. The API base URL comes from --api-url, OVERMIND_API_URL, or overmind.toml; the hosted default is https://api.overmindlab.ai.
overmind init
/overmind slash commands (Cursor and Claude Code only), and the client’s MCP entry, and seeds overmind.toml when it is missing. Other configured MCP servers are left untouched. --env picks the API: production (default), staging, or local (http://localhost:8000); --api-url overrides it. The MCP entry lands in the client’s own file:
When a key is written into that file, the CLI adds it to
.gitignore.
overmind sync
overmind.toml with POST|GET /api/v1/sync. The first sync with an account-scoped key creates the project (named after the repo directory), writes project-id into the toml, mints a project-scoped key into .overmind/credentials.toml, and finishes the MCP configuration. Later syncs push the capabilities /overmind setup wrote and pull reconciled ids back. A capability missing from the local file becomes a leftover on the server; archived = true keeps it one. See Agent & Capabilities.
overmind chassis
/overmind setup treats as ground truth when it writes capability cards. Anchors the coding agent cannot find in the chassis are dropped from the toml before sync.
overmind dataset
upload streams a CSV, TSV, JSON, JSONL, NDJSON, or Parquet file through /api/uploads/ in resumable chunks and lands it as a dataset; the JSON result carries the dataset id, its state, and the MCP calls to make next. --intent is train or eval; omit it and landing proposes one. --split PERCENT lands the file as two datasets, <name> train and <name> eval, with disjoint rows; --split-position (head, tail, or random, default tail) says where the eval rows come from, and the result adds eval_id. --capability binds the dataset on landing.
export downloads one version — the active one unless --cell names another — as JSONL or CSV to a new local file; it refuses to overwrite. The download is a raw stream and never counts as a use. See Datasets.
overmind connector
add TYPE creates a tracing connector — langfuse, langsmith, braintrust, or galileo — from provider credentials it reads from the environment (LANGFUSE_PUBLIC_KEY + LANGFUSE_SECRET_KEY, LANGSMITH_API_KEY, BRAINTRUST_API_KEY, GALILEO_API_KEY, or the generic OVERMIND_CONNECTOR_API_KEY / OVERMIND_CONNECTOR_API_SECRET) or asks for at a terminal prompt; in a non-interactive shell it refuses rather than take them as arguments. --project-id defaults to the project in overmind.toml, --name labels the connector, --base-url (or LANGFUSE_HOST) points at a self-hosted or regional provider API. The connector is created with auto-sync off; the JSON result carries its id and the MCP calls — inspect_connectors, configure_connector, sync_connector — that pick the source project, confirm the capability mapping, and start the import. See Observability.
overmind optimise
start creates the experiment (or attaches to one with -e) and caches the eval dataset version under .overmind/datasets/; next prints the next action for the coding agent. The loop’s verbs, each posting its outputs for the server to score:
start also takes --mode optimize|hybrid, --model (repeatable, hybrid only), --iterations (default 5) and --candidates (default 3). Model comparison runs are scheduled through MCP (start_optimizer) and driven by the same loop.
overmind model
checkpoint.zip from GET /api/deployed-models/{id}/checkpoints/, a presigned link valid for an hour — to a new local file. The deployment id comes from the Console or the overmind://deployments/{deployment} resource. See Inference.
overmind skills
overmind init installs the overmind skill; sync refreshes it in place.
JSON output
--json on dataset upload, dataset export, and model download-checkpoint prints one machine-readable object and nothing else, which is how the MCP prompts hand a local step to the CLI and read the result back.