Skip to main content
rlcli is a lazy-loaded Click CLI that orchestrates local SkyRL-backed Tinker-API post-training on your own GPUs. This page lists every top-level command and the environment variables that control state, paths, and server endpoints so you can find the right tool quickly.

​
Command structure

rlcli uses a lazy-loaded Click group. Subcommands are loaded on demand, so --help on the top level is fast. Run rlcli --help to see the full list, or rlcli <command> --help for a specific subcommand.

​
Top-level commands

​
Environment variables

​
string
default:"~/.rlcli"
Base directory for rlcli state, logs, and the SkyRL source checkout.
​
string
Path to an existing SkyRL checkout. When set, rlcli uses this instead of cloning into ~/.rlcli/skyrl-src.
​
string
default:"http://localhost:8000"
URL of the Tinker API server. Defaults to http://localhost:8000 if unset.
​
string
default:"tml-dummy"
API key for the Tinker SDK. rlcli sets tml-dummy if unset because the SDK requires a value even against local servers that ignore it.
​
string
Fallback tokenizer name or path for rlcli sample when you only pass --checkpoint and no --model.

​
Global flags

  • -h, --help: Show help message for any command or subcommand.

​
Error handling

rlcli catches common errors and prints them as Error: <message> with exit code 1. You will not see a full Python traceback for:
  • LossBackendError: thrown when a chosen loss is not supported by the selected backend.
  • ServerError: thrown on startup failures, timeouts, unknown backends, or duplicate server processes.

​
Next steps

serve

Manage your local SkyRL Tinker server

train

Run supervised, RL, or Harbor agent training

import

Normalize chat dumps to messages JSONL

sample

Quick generation against a checkpoint or base model

Tinker Passthrough

Use checkpoint, run, and session as native tinker commands