ReleaseFireworks AIFireworks AIpublished Oct 5, 2026seen 1w

fw-ai/fireconnect v0.9.8

fw-ai/fireconnect

Open original ↗

Captured source

source ↗
GH
GitHub/github.com/fw-ai/fireconnect
fw-ai/fireconnect v0.9.8
Source ↗
published Oct 5, 2026seen 1wcaptured 1whttp 200method plain

v0.9.8 — Claude Desktop on Fireworks, Claude tier pins, FireRouter OpenAI BYOK

Repository: fw-ai/fireconnect

Tag: v0.9.8

Published: 2026-10-05T00:32:20Z

Prerelease: no

Release notes:

Highlights

  • Claude Desktop runs on Fireworks. On macOS, fireconnect claude-desktop on routes Chat, Cowork, and Code through Fireworks, starting on auto. Bring your connectors along with fireconnect claude-desktop mcp sync, and off puts your previous setup back exactly as it was.
  • Pin Claude Code tiers again. Keep Claude as your main model and hand subagents to an open model: fireconnect claude --sonnet deepseek-flash-latest --subagent deepseek-flash-latest. Re-running on keeps your pins, and native releases a slot.
  • FireRouter can reach OpenAI's frontier models. Save your key once with fireconnect configure --openai-api-key sk-.... It rides along only on FireRouter picks and is never required. Cursor, Copilot, and DeepSeek Harness can now pick plain firerouter too.
  • New Anthropic and OpenAI models show real dollars as soon as models.dev lists them. The status line, fireconnect claude status, and demo estimates read a cached price snapshot, so a model launch no longer waits on a FireConnect release. Offline, the last snapshot still prices your session.

Improvements

  • Fixed: fireconnect claude --model could report success without adding the model. It now becomes your default and shows up in /model, and your own picker rows stay untouched.
  • Fixed: Codex auto-review denied every approval. It called an OpenAI-only reviewer that Fireworks can't serve; now the model you picked does the review.
  • Fixed: --model on a first-time on left Codex with a one-model menu. You now get the model you asked for plus the full catalog.
  • Fixed: DeepSeek Harness 0.1.7 and later asked for a DeepSeek API key on first launch. It now goes straight to Fireworks.
  • Heads-up: Claude Code auto mode may show a "classifier requests are billed" notice once per session. Nothing is broken. Auto mode's safety checks run on your Sonnet model, so they bill to Anthropic unless you pin Sonnet with fireconnect claude --sonnet .
  • New: fireconnect models, a shortcut for fireconnect model list. The list now says when to use each smart router: auto balances cost and quality, auto-instant is latency-first. If your account disables fast models, stay on auto.

> Using Codex auto-review or DeepSeek Harness? Run fireconnect codex on or fireconnect deepseek on once after upgrading to pick up the fixes.

Install

curl -fsSL https://raw.githubusercontent.com/fw-ai/fireconnect/main/install.sh | bash