Skip to main content
FireRouter scores each request to firerouter and routes to either a closed-source model (pass-through) or a Fireworks open model (redirect). The x-routing-preference request header lets you bias that decision toward quality or savings.

​
Preference levels

Send an integer from 1 (most quality-protective) to 5 (most savings-focused):

​
HTTP header

Set the header on each request:
Example with curl:
If the header is missing, invalid, or out of range, FireRouter uses the balanced default (3).

​
FireConnect

When you enable FireRouter through FireConnect, pass --routing-preference on on instead of setting a header yourself:
Supported on Claude Code, OpenCode, Pi, and VS Code. Values are 1–5 or the level names (max-intelligence, balanced, max-savings, etc.). The flag applies when at least one configured slot uses firerouter. Codex, Cursor, and Deep Agents do not support --routing-preference. You can also store a global Anthropic BYOK key once:

​
When to adjust

  • High-volume workloads with many simple requests: try 4 or 5 to redirect summaries, formatting, and straightforward Q&A.
  • Tasks where quality is critical (security review, complex reasoning, nuanced writing): try 1 or 2 to keep closed-source models on harder prompts.
  • Evaluating routing: start at 3 (balanced), then move one step at a time and compare cost and output quality.
FireRouter caches routing decision within a conversation, so preference changes can take a few turns to take effect without restarting your client.