The Hub

The hub is how LLLM packages are shared: validated, published, rendered as cards for people and agents, downloaded, and served. Browse the public hub at lllm.one/hub.

The hub works on lllm.toml packages and comes in three forms:

	What it is	Use it for
Local hub	A folder (`.lllm/hub` by default) managed by `lllm hub`	Publishing and fetching packages on one machine or a shared drive
Hub API	`lllm hub serve`, a FastAPI app over a local hub	Sharing a hub on a network
Static hub	`lllm hub index`, a catalog plus zips any web server can host	Read-only public listings

Validate and publish

lllm hub validate ./briefs
lllm hub publish ./briefs --local          # copies the package into .lllm/hub
lllm hub list

ok
acme/briefs@0.1.0
acme/briefs@0.1.0

Publishing validates first and refuses packages with errors, such as a raw api_key in metadata (see Validation). Secrets and build leftovers (.env, .git, __pycache__, *.egg-info, ...) are never copied. Use --hub DIR (before the command) to work with another hub folder:

lllm hub --hub /shared/lllm-hub publish ./briefs --local

Read a package

lllm hub card acme/briefs             # Markdown card for people
lllm hub agent-card acme/briefs       # operating notes and refs for coding agents
lllm hub config-template acme/briefs  # a starting .lllm/config.toml
lllm hub get acme/briefs --dest ./vendor

Cards are generated from the manifest: description, [card] summary and tags, declared resources with their refs and entrypoints, API-key requirements, and example calls. The agent card is written for coding agents that will use the package: it lists every ref and says how to run the package.

Run a downloaded package

lllm setup ./vendor/briefs        # store the API keys the package declares
lllm serve ./vendor/briefs        # serve its primary resource over FastAPI

lllm serve checks [requirements.api_keys] before starting, reading keys from the environment, a local .env.local, or the OS keyring (the secure extra). See Services for the HTTP API.

Native packages can also be installed as runtime dependencies: a package archive installed with lllm.install_package("briefs.zip") lands in ~/.lllm/packages/ (or ./lllm_packages/ with scope="project"), and lllm.load_runtime() loads it alongside your project. See Discovery.

Serve a hub

lllm hub serve --port 8787

Method and path	Purpose
`GET /health`	Liveness
`POST /validate`	Validate a package folder on the server
`POST /publish`	Publish a package folder on the server
`GET /packages`	List packages
`GET /packages/{org}/{name}`	Package record (`?version=` for a specific one)
`GET /packages/{org}/{name}/card`	Package card
`GET /packages/{org}/{name}/agent-card`	Agent card
`GET /packages/{org}/{name}/config-template`	Local config template
`GET /packages/{org}/{name}/download`	Package zip

lllm.one/hub is its own website, built on the same package format: the lllm-hub repository adds a web app, authenticated uploads, and object storage.

Publish a static hub

lllm hub --hub .lllm/hub index ./site/hub

site/hub/
  data/catalog.json
  packages/acme/briefs/0.1.0.zip

The catalog holds each package's latest version: summary, resources, card, agent card, config template, README, and download link. Like cards and the hub API, it drops secret-shaped metadata keys. Like cards, it also drops them inside JSON schemas, which the hub API returns unchanged (see Hub Metadata). A README that is not valid UTF-8 is read with U+FFFD replacement characters instead of failing the build.

Each zip holds the package files at its root, like GET /packages/{org}/{name}/download, so unzip 0.1.0.zip -d briefs leaves lllm.toml in briefs/, ready for lllm setup briefs and lllm serve briefs.

To make the folder browsable, add the hub web app with lllm-hub --export-web ./site/hub (from the lllm-hub repository); any static host can then serve the folder.

From Python

from lllm.hub import LocalHub, build_catalog, init_package, validate_package

init_package("briefs", org="acme")          # writes briefs/lllm.toml
report = validate_package("briefs")
assert report.ok, report.issues

hub = LocalHub(".lllm/hub")
record = hub.publish("briefs")
print(record.key, [resource.ref for resource in record.resources])
print(hub.card("acme/briefs"))
build_catalog(hub, "site/hub")

lllm.hub.tactic_resource(tactic) and lllm.network.channel_resource(channel) export metadata for tactics and channels you want to describe in a manifest.