Skip to main content
Not ready to rewire your whole setup? fireconnect demo runs the same code-generation prompt through both sides: your current Claude Code provider on Anthropic, and a Fireworks model on the challenger side. It then puts the results next to each other so you can judge speed, cost, and output quality yourself. The demo is read-only against your environment. It uses throwaway config dirs, so your ~/.claude/settings.json is never touched.

​
What you get

  1. Live terminal race: both models stream code in a split-pane TUI with real token counts and wall-clock timers.
  2. Browser comparison: both generated apps run side by side, with measured speed and cost at the bottom.
  3. Shareable artifacts: outputs land in ./fireconnect-demo/ (compare.html, result.json, stream logs).

​
Browser comparison

Side-by-side runnable apps with measured speed and cost:
FireConnect demo browser comparison page with Tetris apps from Anthropic and Fireworks side by side and speed and cost metrics

Browser comparison from a real fireconnect demo run: Anthropic vs Fireworks Tetris apps with measured speed and cost

​
Terminal race

Split-pane TUI while both models stream the same prompt:
Animated capture of a real FireConnect demo split-pane terminal race: Anthropic Claude Sonnet vs Fireworks GLM 5.2 Fast on a Tetris prompt

Real fireconnect demo race: Fireworks GLM 5.2 Fast finishes in 26.3s while Anthropic Claude Sonnet is still running

​
Prerequisites

  • FireConnect installed (overview)
  • Claude Code CLI (claude) on your PATH
  • An Anthropic API key for the incumbent side (ANTHROPIC_API_KEY, or pass --anthropic-key)
  • A Fireworks API key for the challenger side (fireconnect login or FIREWORKS_API_KEY)
Run fireconnect login first so the Fireworks side resolves your key automatically. The demo only supports Claude Code today.

​
Run the demo

That runs the default Tetris preset: both sides get the same prompt to build a playable game in a single HTML file.

​
Try other presets

Presets: tetris (default), snake, tictactoe, clock, or custom with your own task text.

​
Pick the Fireworks model

Default challenger is glm-5p2-fast. Use any serverless model ID you would pass to fireconnect claude on --model.

​
Non-interactive / CI-friendly

​
Clean up

demo clean only removes directories that contain demo markers (result.json, compare.html, etc.), so it will not delete an unrelated folder you pointed --out at by mistake.

​
After the demo

Liked what you saw on the Fireworks side? Wire it into your daily driver:
Want automatic cost routing instead of a fixed model? Try FireRouter:

​
How it works

  • Each side runs real claude -p in an isolated temporary config directory.
  • Numbers in the comparison strip are measured from the run, not list-price estimates.
  • If one side fails to finish, the page says so instead of fabricating a winner.
  • Open compare.html from the output folder anytime. It inlines both apps and works offline.

​
Source

The demo ships with FireConnect: github.com/fw-ai/fireconnect

​
See also