v0.10.0 — FireConnect is now Nexus CLI

Highlights

• FireConnect is now Nexus. The command is nexus (it replaces fireconnect), and your data moves from ~/.fireconnect to ~/.nexus (the old path stays as a link). Your harnesses, key, and settings carry over unchanged. Try nexus status.
• Environment variables are now NEXUS_* (FIRECONNECT_* still works). Rename them in your scripts and shell profiles when you get a chance.
• FireRouter is retired. nexus upgrade moves every harness that was on a FireRouter route to auto, and picking a FireRouter model now fails with a pointer to auto: FireRouter routes are no longer supported. Use --model auto instead.
• Narrow auto to Fireworks-hosted models only: nexus auto --providers fireworks. The setting is global, not per harness. nexus auto --reset restores the default. Claude Desktop and DeepSeek Harness support the setting too.

Improvements

• Claude Desktop works on Linux too: nexus claude-desktop routes Chat, Cowork, and Code through Fireworks there as well as on macOS.

After upgrading, restart your harnesses to pick up the new model settings.

Install

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

v0.9.9 — Codex on FireRouter with your own sign-in, MCP in Codex, key rotation everywhere

Highlights

• Codex starts on FireRouter, and GPT turns run on your own sign-in. A fresh fireconnect codex picks firerouter and asks you to codex login with ChatGPT or an OpenAI key. Easy turns stay on open models, and the GPT models your sign-in offers sit right above the Fireworks catalog in /model. fireconnect upgrade switches existing setups over and keeps your model.
• MCP servers and ChatGPT app connectors work in Codex. Their tools reach the model, so calls complete instead of the model saying a server "exposes no tools". fireconnect upgrade fixes existing setups.
• Rotate your Fireworks key without breaking a harness. fireconnect login now saves the new key into every connected harness, VS Code included, even when you're already signed in. No more 401s from a harness holding the old key.
• OpenCode v2 lists your Fireworks models. When opencode.json uses v2's providers map, FireConnect writes Fireworks there too, so auto and the rest show up in /models instead of silently going missing. fireconnect upgrade repairs existing setups.

Improvements

• Fixed: OpenCode's /models no longer lists routers like glm-latest twice. fireconnect upgrade cleans up existing configs.
• Fixed: Claude Desktop works behind TLS-inspecting corporate proxies such as Zscaler. Re-run fireconnect claude-desktop on to pick it up.
• Fixed: with Node from fnm, fireconnect could stop finding Node in new or non-interactive shells. Re-run the installer once to fix an existing install.
• New: fireconnect configure --unset openai-api-key and --unset anthropic-api-key clear a saved provider key from FireConnect and from every harness holding a copy.
• Clearer errors: when the model catalog can't load, you see the gateway's actual answer instead of a generic "reconnect to the network".
• Heads-up: Claude Code no longer takes --anthropic-api-key. Claude turns already use Claude Code's own sign-in (claude login or an exported ANTHROPIC_API_KEY), and your next fireconnect claude removes any key an earlier version saved.
• Heads-up: Codex no longer supports Azure/Foundry, because Codex CLI 0.160 removed the only transport Foundry offers. Run fireconnect configure --provider fireworks to put Codex back on the Fireworks gateway. OpenCode, Pi, Cursor, and VS Code still support Foundry.

After upgrading, restart Codex and OpenCode. If Codex asks you to sign in, run codex login.

Install

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

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

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 <id> 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 <id>.
• 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

v0.9.7 — GitHub Copilot support, Claude keeps tier models, auto everywhere

Highlights

• GitHub Copilot works with Fireworks. fireconnect copilot-app connects the desktop app, fireconnect copilot-cli the CLI. Your own models stay; off removes only what FireConnect added.
• Claude Code keeps its own tier models. The Fireworks catalog is added to the /model picker instead of overriding slots. Older pinned slots are released on the next fireconnect claude on.
• Every other harness starts on auto, the cost-aware mix, also selectable in every picker.
• Claude Code web search and fetch go through Fireworks natively; the old fireworks-websearch MCP is removed on upgrade, your own MCPs untouched.
• Codex web search works again; upgrade cleans up the 0.9.6 override.

Improvements

• The built-in Explore agent follows your session model again instead of always using Opus. Fixed on upgrade.
• Pickers stay fresh: re-running on prunes retired models, adds new ones, and updates pricing and metadata. Upgrade refreshes the Claude picker too.
• Codex auto-compacts at 80% of the context window so long sessions survive; GLM-5.3 models unlock max reasoning effort.

Install

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

v0.9.6 — Codex 0.153 fix, cleaner cost breakdown

Highlights

• Codex 0.153 works with Fireworks again. Codex 0.153 turns on a web_search tool the Responses API can't accept; FireConnect now bakes web_search = "disabled" and repairs existing setups on fireconnect codex on.
• Cleaner cost breakdown — the status line and meter dedupe and name model labels instead of listing the same one twice.

Improvements

• Claude Fable 5.1 pricing added; estimates and fireconnect model list are accurate.
• Subagent slot keeps its [1m] tag when routed.
• fireconnect model list aligns price precision across models.
• fireconnect claude demo disables your user MCPs so they can't skew the race.

Using Codex? Run fireconnect upgrade and 0.153 keeps working through Fireworks.

Install

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

v0.9.5 — Claude status line, ChatGPT desktop app, smart router mixes (preview)

Highlights

• Session costs in your Claude prompt. A new status line shows which models ran, what they cost at Fireworks rates, and cache savings. Your own status line is left alone — fireconnect claude off restores everything.
• ChatGPT desktop app support. fireconnect chatgpt and fireconnect codex share one config, so you route both through Fireworks with a single command. Quit the app first so its model list refreshes.
• Smart router mixes (preview). Pin auto for the default mix or auto-instant for latency-first routing — for example, fireconnect claude --sonnet auto-instant. Every auto-* slug works the same way.
• Simpler Claude model setup. The wizard lets you pick a model per slot instead of toggling fast profiles. New installs default to Sonnet on deepseek-pro-latest, Opus on glm-latest, and Fable on vision-capable glm-flash-latest. FireRouter still lands on Opus at first connect when your key supports it.
• Faster fireconnect model list. Results are cached for an hour, work offline, show when they were last updated, and --refresh pulls the latest. Smart Routers lists auto, auto-instant, and FireRouter.

Improvements

• MCP tool search works through the Fireworks gateway again — upgrade applies the fix automatically.
• fireconnect claude demo and the status line use the same pricing, so their numbers always match.
• fireconnect upgrade repairs missing dependencies; partial installs recover instead of crashing.
• Linux SSH/WSL: keys store in an encrypted file with timeouts on keyring and network calls — no more hangs.
• fireconnect claude status shows every slot, including defaults you never explicitly set.
• GLM 5.3 and GLM 5.3 Flash added. Stale pins no longer hide newer releases; US-only pricing corrected.
• Claude picker lists auto on every slot and auto-instant on Sonnet only.

Already on Claude? Upgrade keeps your slot picks — run fireconnect claude --interactive to adopt the new defaults. Legacy pins like deepseek-v4-flash move to deepseek-flash-latest on your next fireconnect claude on.

Install

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

v0.9.4 — Claude demo cost accuracy, safer Cursor on/off and uninstall

Highlights

• fireconnect claude demo: improved cost calculation and racing experience.
• fireconnect cursor on/off: more robust; uninstall now waits for Cursor to quit before restoring, preventing broken state.
• fireconnect uninstall: paced, harness-by-harness checklist that waits for IDEs to quit.

Install

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

v0.9.3 — Anthropic+Fireworks Claude defaults, DeepSeek harness, demo

Highlights

• New Claude Code defaults: Sonnet on Anthropic's own model, Opus on deepseek-pro-latest, Haiku and subagents on deepseek-flash-latest. Change them with fireconnect claude --interactive.
• Anthropic and Fireworks models both work once fireconnected — mix them freely.
• fireconnect claude demo races two live Claude Code sessions side by side.
• DeepSeek Harness replaces Deep Agents — fireconnect deepseek on.

Improvements

• fireconnect status no longer reports "All off" while models are routed.
• OpenCode shows per-model cost, and hides it for Fire Pass subscriptions.

Pinned deepseek-v4-flash? Your next fireconnect claude on moves it to deepseek-flash-latest.

Install

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

v0.9.2 — Claude live usage meter, Cursor/VS Code fixes, harness improvements

Highlights

• Live session cost meter — watch Fireworks spend accumulate in a tmux split pane beside Claude Code, with a recent-session picker.
• VS Code now uses the chat-completions API for better compatibility.
• Cursor picker hides built-in models and preserves your native model IDs when re-enabling.

Improvements

• FireRouter model IDs are treated as 1M-context; Kimi K3 specs corrected to 1M with [1m] enabled for Claude.
• Linux quit detection ignores Electron helper processes in VS Code.
• Claude usage meter exits cleanly on quit.
• README rewrite.

Install

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

FireConnect v0.9.1

FireConnect v0.9.1

Claude harness

• Interactive model onboarding with persisted preferences (#224)
• Fast model profile toggle (#232)
• Subscription model picker uses catalog labels (#207)
• Kimi/GLM slot defaults, [1m] for Kimi K3, main → kimi-fast-latest (#218)
• Store main default in settings.model, not ANTHROPIC_MODEL (#228)
• Bake websearch MCP Bearer token like claude mcp add (#234)
• Remove unused SessionStart desktop-guard hook (#233)
• Warn on text-only models and document vision limits (#204)

Kimi K3 pricing & display

• Correct Kimi K3 -latest alias pricing tiers (#215)
• Correct Kimi K3 pricing and hide priority tier from catalog (#217)
• Show Kimi K3 for kimi-latest router display (#214)
• Fix fast-latest pricing fallback when catalog lacks API rates (#209)

FireRouter

• Require Anthropic creds or workspace BYOK before explicit on (#198)
• Improve on output and enable Cursor with workspace BYOK (#194)
• Echo routing preference only when explicitly set (#193)
• Remove duplicate override message from status (#192)

OpenCode

• Set context limits on router alias model overrides (#223)
• Dedupe provider.models keys on repeat on (#216)
• Fix catalog model handling and kimi vision modalities (#201)

Install & upgrade

• Skip Claude-off on 0.9.0+; prompt to upgrade (#235)
• Fix install.sh crash when INSTALL_NOTES is empty under set -u (#206)

Other harnesses

• Cursor: cache macOS Safe Storage keychain reads (#225), Linux libsecret detection (#205)
• VS Code: tighten Linux pgrep pattern (#222)
• Codex: exclude minimax models from catalog (#203)

CLI & docs

• Fold environment info into status, remove env command (#187)
• Rewrite README (#227)
• Document safe upgrades from 0.7 and 0.8 (#191)