Skip to main content
What changed in each Nexus CLI release, newest first. To update, run:
Restart your harnesses afterward so they load the new settings. Run nexus --version to see which version you have. For every deprecated or removed command in one table, see Deprecated and removed.
Versions before 0.10.0 shipped as FireConnect, with the fireconnect command, and Nexus Auto was called FireRouter. Their notes use today’s command names. See Upgrading from FireConnect.
​
October 22, 2026
New
  • Bring your own Anthropic and OpenAI keys to Nexus Auto. Run nexus auto --anthropic-api-key or nexus auto --openai-api-key. Each prompts for the key, so it stays out of your shell history. Then run your harness command again. See Nexus Auto.
  • Every key in one place. nexus auto shows the keys saved on this machine next to the Provider Keys on your Fireworks account. Provider Keys work in every harness, and a key saved on your machine wins for its provider.
  • nexus status lists installed harnesses, not just connected ones.
  • nexus update updates Nexus CLI. nexus upgrade still works.
Improved
  • Connecting a harness that isn’t installed now stops with an explanation and writes nothing. Pass --skip-install-check if the tool lives somewhere Nexus CLI doesn’t look.
  • Claude Code: the Artifact tool shows up, a Claude Code older than 2.1.242 gets a warning instead of an empty /model picker, and the status line names the model that served the latest turn. Run nexus claude again to refresh the picker.
  • Fewer keychain prompts. Nexus CLI reads your keychain only when a command needs your key.
  • Sign-in no longer tells account members that they aren’t members.
  • nexus logout also clears the Anthropic and OpenAI keys saved on this machine.
  • Safer updates: if ~/.nexus wasn’t created by Nexus CLI, nexus update stops before changing anything.
  • The installer now requires Node.js 22 or later.
Deprecated and removed
  • Removed nexus claude live. The Claude Code status line shows the model and spend on every turn, and nexus claude usage gives the full breakdown.
  • Removed --anthropic-api-key and --openai-api-key from harness commands such as nexus opencode. Save keys with nexus auto instead.
  • Deprecated nexus harnesses. Use nexus status.
  • Deprecated --anthropic-api-key, --openai-api-key, and --unset on nexus configure. Use the same flags on nexus auto.
​
October 20, 2026
New
  • FireConnect is now Nexus CLI. The command is nexus, and your data moves from ~/.fireconnect to ~/.nexus. The old path stays as a link. Your harnesses, key, and settings carry over. Try nexus status.
  • FireRouter is now Nexus Auto. Use auto everywhere. The update moves every harness on a FireRouter route to auto.
  • Keep Nexus Auto on Fireworks models with nexus auto --providers fireworks. The setting is global, and nexus auto --reset restores the default. See Nexus Auto.
  • Claude Desktop runs on Linux as well as macOS. See Claude Desktop.
Deprecated and removed
  • Removed FireRouter models (firerouter, firerouter/...) and the auto-instant mix. Picking one fails and points you to auto.
  • Deprecated FIRECONNECT_* environment variables. They still work. Rename them to NEXUS_* in your scripts and shell profile.
  • Deprecated --routing-preference. It has no effect. Set effort in your harness instead.
​
October 12, 2026
New
  • Codex starts on FireRouter (now Nexus Auto), and OpenAI turns use your own sign-in. Run codex login with ChatGPT or an OpenAI key. Easy turns stay on open models, and the OpenAI models your sign-in offers sit above the Fireworks catalog in /model. See Codex.
  • MCP servers and ChatGPT app connectors work in Codex. Their tools now reach the model.
  • Rotate your Fireworks key without breaking a harness. nexus login saves the new key into every connected harness, VS Code included, even when you’re already signed in.
  • OpenCode v2 lists your Fireworks models. When opencode.json uses the v2 providers map, Nexus CLI writes Fireworks there too.
Fixed
  • OpenCode’s /models no longer lists models such as glm-latest twice.
  • Claude Desktop works behind TLS-inspecting corporate proxies such as Zscaler. Run nexus claude-desktop on again to pick it up.
  • With Node.js from fnm, the CLI could stop finding Node in new shells. Re-run the installer once to fix it.
  • When the model catalog can’t load, you see the gateway’s actual error.
Deprecated and removed
  • Removed --anthropic-api-key from nexus claude. Anthropic turns in Claude Code use Claude Code’s own sign-in.
  • Removed Codex on Microsoft Foundry, because Codex CLI 0.160 dropped the only API Foundry offers. Run nexus configure --provider fireworks, then nexus codex. OpenCode, Pi, Cursor IDE, and VS Code still support Foundry.
​
October 4, 2026
New
  • Claude Desktop runs on Fireworks. On macOS, nexus claude-desktop on sends Chat, Cowork, and Code through Fireworks, starting on auto. off puts your previous setup back exactly as it was.
  • Pin Claude Code tiers. Keep Claude as your main model and hand subagents to an open model: nexus claude --sonnet deepseek-flash-latest --subagent deepseek-flash-latest. Pass native to release a slot.
  • FireRouter (now Nexus Auto) can use OpenAI models with your own OpenAI key. Cursor IDE, the Copilot harnesses, and DeepSeek Harness can pick it too.
  • New Anthropic and OpenAI models show real prices as soon as models.dev lists them. Prices come from a cached snapshot, so session estimates still work offline.
  • nexus models is a shortcut for nexus model list.
Fixed
  • nexus claude --model <id> now makes the model your default and adds it to /model.
  • Codex auto-review runs on the model you picked, instead of denying every approval.
  • --model on a first Codex connection keeps the full catalog in the picker.
  • DeepSeek Harness 0.1.7 and later go straight to Fireworks, with no DeepSeek key prompt.
Using Codex auto-review or DeepSeek Harness? Run nexus codex or nexus deepseek once after updating.
​
September 21, 2026
New
  • GitHub Copilot works with Fireworks. nexus copilot-app connects the desktop app, and nexus copilot-cli connects the CLI. Your own models stay, and off removes only what Nexus CLI added.
  • Claude Code keeps its own tier models. The Fireworks catalog joins the /model picker instead of replacing slots.
  • Every other harness starts on auto.
  • Native web search and fetch in Claude Code, served by Fireworks. See Web Search.
  • Codex web search works again.
Improved
  • The built-in Explore agent follows your session model again, instead of always using Opus.
  • Connecting again refreshes the picker: retired models go, new ones arrive, and pricing updates.
  • Codex compacts at 80% of the context window, so long sessions keep going.
  • GLM 5.3 models unlock the max reasoning effort.
Removed
  • The fireworks-websearch MCP server. The update removes it and leaves your own MCP servers alone.
​
September 3, 2026
Fixed
  • Codex 0.153 works with Fireworks again.
  • The status line and session meter name each model once, instead of listing it twice.
  • Claude Fable 5.1 pricing is in estimates and nexus model list.
  • A routed subagent slot keeps its [1m] tag.
  • nexus model list lines up prices across models.
  • nexus claude demo turns off your own MCP servers so they can’t skew the race.
​
August 31, 2026
New
  • Session cost in Claude Code. A status line shows which models ran, what they cost at Fireworks rates, and your cache savings. If you already have a status line, Nexus CLI leaves it alone. See Usage and Cost.
  • ChatGPT desktop app support. nexus chatgpt and nexus codex share one config, so one command connects both. See ChatGPT Desktop.
  • Pick a model for each Claude Code slot in the setup wizard.
  • Faster nexus model list. Results are cached for an hour, work offline, and show when they were last updated. --refresh pulls the latest.
Improved
  • MCP tool search works through the Fireworks gateway again.
  • nexus claude demo and the status line use the same prices, so their numbers match.
  • The update repairs missing dependencies, so a partial install recovers.
  • On Linux over SSH or WSL, keys go to an encrypted file, with timeouts on keyring and network calls.
  • nexus claude status shows every slot, including defaults you never set.
  • GLM 5.3 and GLM 5.3 Flash are in the catalog.
​
August 20, 2026
Improved
  • nexus claude demo estimates cost more accurately.
  • Connecting and disconnecting Cursor IDE is more reliable.
  • nexus uninstall restores one harness at a time and waits for open IDEs to quit.
​
August 18, 2026
New
  • Anthropic and Fireworks models side by side in Claude Code. Mix them freely.
  • nexus claude demo races two live Claude Code sessions side by side.
  • DeepSeek Harness replaces Deep Agents: nexus deepseek. See DeepSeek Harness.
Fixed
  • nexus status no longer reports “All off” while models are connected.
  • OpenCode shows the cost of each model.
Deprecated
  • nexus demo. Use nexus claude demo.
​
August 9, 2026
New
  • VS Code uses the Chat Completions API for better compatibility.
  • Cursor IDE’s picker hides the built-in models while connected.
  • A live session cost meter for Claude Code. Today it opens with nexus claude usage.
Fixed
  • Kimi K3 lists its 1M-token context window.
  • On Linux, quit detection ignores VS Code’s Electron helper processes.
​
August 1, 2026
New
  • Claude Code stays connected when you reinstall or update.
  • Update in place. When a newer version is out, interactive terminals offer to update. Answer No to snooze for a day.
  • Claude Code model choices persist across sessions, with a guided setup.
Improved
  • More reliable setup for Cursor IDE, OpenCode, and VS Code.
  • Clearer Claude Code model defaults and catalog labels, including Kimi K3 naming and pricing.
​
July 26, 2026
New
  • nexus model list shows the coding model catalog in one place, with standard and fast options across model families.
  • FireRouter (now Nexus Auto) works like any other model through --model.
  • Claude Code usage reports show session tokens and estimated Fireworks cost.
Improved
  • Sign-in supports enterprise SSO and stores your key securely.
  • Your own settings, headers, and model selections are kept when you connect and disconnect.
  • More reliable Microsoft Foundry endpoints, Windows installs, and error messages.