Hermes Agent v0.14.0 lands with a change that matters immediately for developers running AI coding tools: a new OpenAI-compatible local proxy. Point Codex, Aider, Cline, or Continue at it, and it routes requests through whichever OAuth-authenticated provider you already pay for, whether that is Claude Pro, ChatGPT Pro, or the new SuperGrok integration. No extra API key. No separate billing account. You reuse the subscription you already have.
The SuperGrok OAuth provider is the headline addition on the model side. If you pay for SuperGrok, you authenticate with your xAI account and get Grok inside Hermes. The release also bumps grok-4.3 to a 1M token context window, which means you can send an entire codebase or research corpus in a single prompt. The team added proper entitlement error handling and a dedicated SSH-to-tunnel docs page for developers completing the OAuth flow from a remote machine.
The release is built around what the team calls a Foundation Release philosophy: install anywhere, ship what you need, drop what you do not. pip install hermes-agent now works from PyPI with no friction. Heavyweight backends lazy-install on first use rather than at install time. The [all] extras no longer pull in dependencies covered by lazy installs. A tiered install falls back gracefully when a wheel rejects on your platform. The result is a dramatically lighter install footprint.
Performance numbers are concrete. Cold-start time drops by roughly 19 seconds. Browser CDP calls are 180x faster. Those two changes alone shift Hermes from feeling like a heavy tool to feeling like a responsive one.
The release covers a lot of additional ground. Two new messaging platforms, LINE and SimpleX Chat, bring the total to 22. Microsoft Teams is now wired end-to-end with Graph auth, webhook listener, pipeline runtime, and outbound delivery. Cross-session Claude prompt caching runs at a 1-hour window. A /handoff command transfers sessions live. The computer_use cua-driver backend now works with non-Anthropic providers. LSP semantic diagnostics fire on every write. A unified pluggable video_generate interface and 9 new optional skills round out the feature additions.
The full changelog covers 808 commits, 633 merged PRs, and 545 closed issues including 12 P0 and 50 P1 fixes, with 215 community contributors involved.
What should you do today: if you are running Aider or Cline against an OpenAI API key just to get a compatible endpoint, swap in the Hermes local proxy and point it at your existing Claude Pro or ChatGPT Pro account. You get the same interface, drop the per-token API cost for that workflow, and pick up the 1M context window if you are on SuperGrok.