CLI
lllm is the single command. Command groups (hub, and the plugin commands
network and shell) take their own options before the subcommand.
lllm create NAME [--template T] new project
lllm init [PATH] write lllm.toml
lllm setup [TARGET] store the API keys a package declares
lllm inspect TARGET tactic metadata, or what serve launches
lllm serve TARGET serve a tactic or a package resource
lllm hub ... validate, publish, read, index packages
lllm network ... SSSN channel stores (pip install sssn)
lllm shell ... AAAX strategies (pip install aaax)
TARGET is a module:attribute entrypoint (imported from the current
directory) or a package path (a folder with lllm.toml, or the file itself).
lllm create
lllm create my-app # template "minimal"
lllm create my-app --template research # or pipeline, proxy, service
lllm create my-svc --template pydantic-ai # a typed service scaffold
lllm create my-app --template ./my-template # has lllm-template.toml
The templates minimal, pipeline, proxy, research, and service
create native projects; plain, pydantic-ai, and native create typed
service scaffolds. The older form lllm create TEMPLATE NAME also works.
| Option | Default | |
|---|---|---|
--template |
minimal |
Template name or folder |
--model |
gpt-4o-mini |
Model written into configs/default.yaml |
--directory |
. |
Parent folder |
--force |
off | Overwrite existing files |
Native templates create lllm.toml, prompts/, configs/default.yaml,
tactics/, main.py (with create_tactic()), and an offline smoke test.
lllm init
lllm init [PATH] [--name NAME] [--org ORG] [--kind KIND] [--force]
Writes a minimal lllm.toml (with [card] and [docs.readme]) and a README
if there is none. See The manifest.
lllm setup
lllm setup ./briefs # prompts for missing keys
lllm setup ./briefs --credentials env --env-file .env.local \
--set OPENAI_API_KEY=sk-...
lllm setup --show-conventions # common provider key names
Reads [requirements.api_keys], plus the required_api_keys metadata of the
resource lllm serve would launch (or the one named with --resource), and
stores missing keys in the OS keyring (secure extra) or a local env file. A
package with nothing to launch, such as a library, is set up from its
[requirements.api_keys] alone. --credentials auto uses the keyring when
available.
lllm inspect
lllm inspect tactics:Shout --json # name, schemas, capabilities, endpoints
lllm inspect ./briefs # resource, port, routes, key status
Package options: --resource section.name, --store DIR (channel
resources), --env-file, --no-keyring.
lllm serve
lllm serve tactics:Shout --port 8000
lllm serve ./briefs # the package's primary resource
lllm serve ./briefs --resource services.api
| Option | Default | |
|---|---|---|
--host |
127.0.0.1 |
Bind address (host or IP, not a URL) |
--port |
declared port or 8000 | |
--log-level |
info |
Uvicorn log level |
--resource |
[package] primary |
Resource to launch |
--store |
<package>/.lllm/network |
Store root for channel resources (needs sssn) |
--credentials, --env-file, --no-keyring |
Key lookup and interactive setup | |
--skip-key-check |
off | Start without checking declared keys |
An entrypoint names a tactic class or instance, a factory, an object with a
run method, or a handler function. A resource that cannot be launched, or a
service entry that fails to import or build, stops with an error that names
the cause. See Serving.
lllm hub
lllm hub [--hub DIR] COMMAND
--hub defaults to .lllm/hub.
| Command | |
|---|---|
init [PATH] |
Same as lllm init |
validate PATH |
Validate a package |
publish PATH --local [--no-validate] |
Publish into the local hub |
list |
Published packages (org/name@version) |
get ID [--dest DIR] [--version V] |
Download a package |
card ID, agent-card ID |
Render the package card / agent card |
config-template ID |
Render a starting .lllm/config.toml |
index OUTPUT [--generated-at TS] |
Write a static catalog for the hub web app |
serve [--host] [--port 8787] |
Serve the hub API |
See The Hub.
Plugin commands
Other packages add commands to lllm through the lllm.commands entry-point
group, which maps a command name to a module with main(argv):
[project.entry-points."lllm.commands"]
network = "sssn.cli"
lllm --help lists the installed ones. A plugin's main(argv) runs as it
would under the package's own command: lllm imports it from the installed
package and does not put the working directory on sys.path, so a local
sssn.py or aaax.py cannot shadow it. LLLM's companion packages add two:
| Command | Install | Docs |
|---|---|---|
lllm network [--store DIR] COMMAND |
pip install sssn or "lllm-core[network]" |
sssn.one |
lllm shell inspect TARGET, lllm shell serve TARGET |
pip install aaax or "lllm-core[shell]" |
aaax.one |
Each package also keeps its own command, sssn or aaax. Without the
package, the lllm command says what to install and exits with status 2:
lllm network needs the sssn package: pip install sssn