Skip to main content
The Fireworks training skill teaches your coding agent to run training end to end: choose managed or Training API, choose serverless or dedicated infrastructure when needed, prepare and validate data, launch and monitor the run, evaluate, deploy, and troubleshoot.

​
Install

​
Claude Code

Install the auto-updating cookbook plugin:

​
Cursor

​
Codex

​
Other compatible agents

Install to every detected Agent Skills-compatible harness:
Skills installed with npx skills do not update automatically. Refresh them with npx --yes skills update -g -y.

​
Prerequisites

  • firectl installed. Authenticate with either firectl signin or FIREWORKS_API_KEY.
  • Export FIREWORKS_API_KEY for Training API Python workflows.

​
What it does

Ask in plain language, for example “Fine-tune qwen3-4b on my train.jsonl and deploy it.” The skill uses current Fireworks docs and cookbook recipes to choose the right path, validate inputs, run and monitor training, evaluate the result, deploy it, and troubleshoot failures. Before any upload, registration, paid inference, job creation, promotion, or deployment, it shows the resolved parameters, defaults, cost ceiling, evaluation, and teardown for approval. Material changes require approval again; promotion and deployment each have separate confirmation.
firectl may block mutating commands inside any AI-agent environment. When that happens, the skill gives you the exact command to run manually in your terminal, then resumes read-only monitoring and reporting. It does not bypass the guard.
Use managed fine-tuning for standard jobs, or the Training API for custom loops on serverless or dedicated infrastructure.

​
See also

Managed Fine-Tuning

Drive the same training infra directly when you know your config.

Training API

Write your own Python training loop on Fireworks GPUs.

Choose infrastructure

Compare serverless and dedicated training.

firectl

The CLI the skill drives.

Cookbook

Ready-to-run recipes, including the inline-reward RL loop.