kitops

name	kitops
description	Package, version, share, and unpack AI/ML models as OCI-standard ModelKits with the KitOps `kit` CLI. Use this skill when authoring a Kitfile or running kit init, pack, tag, login, push, pull, or unpack.
license	Apache-2.0

Using the KitOps kit CLI

KitOps packages models, datasets, code, docs, prompts, and MCP servers together as a ModelKit — an OCI artifact you can store in any OCI registry. A ModelKit is defined by a Kitfile (a YAML manifest) and built with kit pack.

This skill documents the real behavior of the kit CLI. It was generated from and ships with a specific kit version, so it always matches the installed binary. Confirm anything not covered here with kit <command> --help.

Core workflow

kit init <dir>     # generate a Kitfile for a directory (optional starting point)
kit pack <dir>     # build a ModelKit into local storage
kit tag SRC DST    # add another reference to a local ModelKit
kit login <reg>    # authenticate to a remote registry
kit push SRC       # upload a ModelKit to a registry
kit pull REF       # download a ModelKit into local storage
kit unpack REF     # extract a ModelKit's contents onto disk

Authoring a Kitfile

The Kitfile is a YAML manifest, by default named Kitfile at the root of the context directory. All path values are relative to the context directory; absolute paths are rejected. Only local filesystem paths are allowed for code, datasets, docs, prompts, and mcpServers paths; model.path also accepts a ModelKit reference. No two layers may declare the same path.

Sections:

• manifestVersion (string): use 1.0.0. Other values produce a warning and are treated as 1.0.0.
• package: name, version, description, authors (list of strings).
• model (object): name, path, framework, format, version, description, license (SPDX id), parts (list of name/path/type, e.g. LoRA weights), and parameters (arbitrary JSON-compatible YAML).
• code (list): each has path, description, license.
• datasets (list): each has name, path, description, license. A dataset may instead reference a remote source via remotePath (an s3:// URL, which also requires remoteHash, or a ModelKit reference).
• docs (list): each has path, description.
• prompts (list): each has path, description. A prompt directory that contains a SKILL.md can be installed as an agent skill (see unpack).
• mcpServers (list): each has name (required, unique) and path pointing to a single .mcpb bundle file, plus description.

Example:

manifestVersion: 1.0.0
package:
  name: mymodel
  version: 1.0.0
  description: Sentiment classifier
  authors: [Jane Doe]
code:
  - path: src/
    description: Training and inference code
    license: Apache-2.0
datasets:
  - name: training-set
    path: data/train.csv
    description: Labeled training data
    license: CC-BY-4.0
model:
  name: sentiment-model
  path: models/model.safetensors
  framework: PyTorch
  version: 1.0.0
  license: Apache-2.0

Tip: kit init <dir> inspects a directory and generates a starting Kitfile. Useful flags: --name, --desc, --author, --force (overwrite an existing Kitfile), and --output (- writes to stdout). kit init <repo> --remote generates a Kitfile from a remote Hugging Face repository (with --ref).

pack — build a ModelKit

kit pack [flags] DIRECTORY

Builds a ModelKit from the Kitfile using DIRECTORY as the context, and stores it in local storage. Relative paths in the Kitfile are resolved against DIRECTORY.

• -f, --file — path to the Kitfile (default: Kitfile in the context dir; use - to read from stdin).
• -t, --tag — assign one or more tags, comma-separated: -t registry/repository:tag1,tag2.
• --compression — none (default), gzip, gzip-fastest, or zstd.
• --layer-format — tar (default) or raw.

kit pack .
kit pack . -f ./Kitfile -t myregistry.com/myorg/mymodel:1.0.0

tag — add a reference

kit tag SOURCE_MODELKIT[:TAG] TARGET_MODELKIT[:TAG]

Adds a new reference to an existing local ModelKit. Both the source and target must include a tag or digest.

A full reference is [HOST[:PORT]/][NAMESPACE/]REPOSITORY[:TAG]. HOST defaults to localhost when omitted. A TAG may use letters, digits, _, ., -, must not start with . or -, and is at most 128 characters.

kit tag myregistry.com/myorg/mymodel:latest myregistry.com/myorg/mymodel:v1.0.0

login — authenticate

kit login [flags] [REGISTRY]

Authenticate before pushing or pulling from a private registry.

• -u, --username, -p, --password, --password-stdin.

kit login myregistry.com -u myuser --password-stdin

push — upload to a registry

kit push [flags] SOURCE [DESTINATION]

Uploads a local ModelKit to a remote registry. Without a DESTINATION, the ModelKit must already be tagged with a full registry reference (a registry host is required — you cannot push to localhost). With a DESTINATION, a locally tagged ModelKit is pushed to that reference.

kit push myregistry.com/myorg/mymodel:1.0.0
kit push mymodel:1.0.0 myregistry.com/myorg/mymodel:latest

pull — download from a registry

kit pull [flags] registry/repository[:tag|@digest]

Downloads a ModelKit into local storage. A registry host is required (you cannot pull from localhost).

kit pull myregistry.com/myorg/mymodel:1.0.0

unpack — extract contents to disk

kit unpack [flags] [registry/]repository[:tag|@digest]

Extracts a ModelKit's contents to the filesystem. Looks in local storage first, then the remote registry.

• -d, --dir — target directory (created if missing; default: current dir).
• -f, --filter — limit what is extracted, format [types]:[filters], where types is a comma-separated subset of kitfile,model,datasets,code,docs,prompts,mcpservers. Repeatable; a layer is unpacked if it matches any filter. Example: --filter=model or --filter=datasets:my-dataset.
• -o, --overwrite — overwrite existing files.
• -i, --ignore-existing — skip files that already exist.
• --as-skill[=agent,...] — install SKILL.md prompt layers as agent skills instead of unpacking to their original paths. With no value, kit auto-detects installed agents; otherwise pass a comma-separated list (e.g. --as-skill=claude-code,cursor). Installs globally unless -d is given.

kit unpack myregistry.com/myorg/mymodel:1.0.0 -d ./unpacked
kit unpack myregistry.com/myorg/mymodel:1.0.0 --filter=model -o

Other useful commands

• kit list — list ModelKits in local storage (or a remote repository).
• kit inspect REF — print a ModelKit's manifest and config.
• kit info REF — show a ModelKit's Kitfile.
• kit remove REF — delete a ModelKit from local storage.
• kit version — print the installed kit version.

Common errors and how to resolve them

• "registry is required when pushing" / "...when pulling" — the reference has no registry host (it resolves to localhost). Use a full reference like myregistry.com/myorg/mymodel:tag, and kit tag the local ModelKit first if needed.
• "No tag specified ... Using 'latest' as default" — informational; kit defaulted the reference to :latest. Specify a tag explicitly to avoid it.
• "reference cannot include multiple tags" — push/pull/unpack accept a single reference. Split multiple tags into separate commands (or use kit tag).
• "source/target ModelKit reference requires a tag or digest" — kit tag needs a :tag (or @digest) on both arguments.
• Kitfile validation errors (surface on pack/init) include:
  • "absolute paths are not supported in a Kitfile" — make every path relative to the context directory.
  • "only local paths are permitted" — code/datasets/docs/prompts paths must be local files or directories, not URLs.
  • "... use the same path ..." — two layers point at the same path; each layer needs a distinct path.
  • "invalid path for mcpServer ...: path must point to a single .mcpb file" — an mcpServers entry must reference one .mcpb bundle.
  • "Unrecognized manifestVersion" — set manifestVersion: 1.0.0.
• "Failed to pack model kit" — a path in the Kitfile is missing or the Kitfile itself was not found; verify -f/the context directory and that every path exists.