Hermes Agent ships Herald with Ironproxy secrets lockdown

• Herald is a runtime-shaped Hermes Agent release, not just a model-provider bump: Teknium's release inventory lists voice chats, desktop plugins, Kanban, Agent2Agent, outbound webhooks, grounded research, Buzz, Ironproxy, and productivity skills.
• The densest engineering work is the tool-efficiency batch: Teknium's performance post says Hermes mined 250,000 conversations, used NVIDIA NeMo Relay, and shipped changes aimed at wasted turns, tool errors, schemas, and token load.
• Voice moved from a feature to a surface: Teknium's provider reply says Hermes can hook into major voice providers including ElevenLabs, while Teknium's microphone reply reduced the hardware requirement to a microphone.
• Secrets handling got a real boundary: Teknium's release inventory includes Ironproxy lockdown, and the Hermes Ironproxy docs describe sandboxes holding opaque proxy tokens instead of real API keys.
• The open-model path is getting louder: Teknium's Qwen post says Qwen 3.8 Max and a local 27B Qwen 3.8 are coming to Hermes Agent, and Alibaba_Qwen's reply points users toward Qwen3.8-Max on Hermes.

Christmas came early for coding-agent runtime nerds: the v0.20.0 GitHub release claims roughly 3,650 commits, 1,400 merged PRs, 5,200 changed files, and 1,200 closed issues since v0.19.0. The voice docs now describe sentence-by-sentence TTS and full-duplex barge-in. The grounded-citations docs reject quotes unless they literally appear in fetched page text.

The official release tag is Hermes Agent v0.20.0, v2026.8.3, published August 3, according to the GitHub release notes. Teknium's X post compressed the release into the practical feature list engineers will scan first:

• Voice activated chats
• Plugins and Kanban for the desktop GUI app
• Agent2Agent protocol
• Outbound webhooks
• Performance optimizations and token-efficiency work
• Buzz integration
• 20% discount on all models in Nous Portal
• More reliable local and open models
• Office and productivity skills
• Grounded research skill
• Secrets lockdown with Ironproxy

The shipping style stayed Teknium-coded: Teknium's teaser posted “Stop, Hermes Time” before the larger release thread.

Hermes voice now spans CLI, desktop, and audio-capable gateways, according to the v0.20.0 release notes. The voice mode docs say TTS buffers model output into sentences, strips markdown and code formatting, then plays audio in real time.

Barge-in is the bigger behavioral change: the docs say the mic stays live during the agent turn, so an interruption can land while Hermes is still thinking or speaking. Teknium told one user that Hermes hooks into major voice providers, including ElevenLabs, in a release-thread reply.

Release-day replies exposed the practical surface area:

• Voice input needs “a microphone,” per Teknium's microphone reply.
• Slack and Discord already work, according to Teknium's gateway reply.
• Photon can be used to set up iMessage, per Teknium's Photon note.
• One asked-about path existed but was “not well tested,” according to Teknium's testing caveat.

Teknium framed the performance work as trace-driven engineering across 250,000 Hermes conversations. The GitHub tracker says the performance target was “fewer wasted tool turns, fewer model-confusing errors, smaller/clearer schemas,” not only wall clock.

The attached batch chart breaks the work into five evaluation steps:

1. Mine 1.53M messages for error classes and wasted turns.
2. Survey 13 agent codebases plus research synthesis.
3. Add fixes, unit tests, and live E2E checks per PR.
4. Run NVIDIA NeMo Relay ATOF benchmark batteries.
5. Run hard A/B evals across 2 models, 9 error-inducing tasks, and 3 repetitions.

The same chart lists 16 improvements across terminal, file, patch, search, process, skills, and sandbox. The headline numbers are the kind of runtime work that makes weaker/local models feel less cursed:

• Weak model: 21% fewer LLM turns.
• Tool calls: 29% fewer.
• Tool errors: reduced to 0 in the eval batch.
• Wall clock: 23% lower.
• Schema diet: about 700 tokens saved per request.
• Repeat skill view dedup: about 24.8K tokens saved.
• Truncated reads: 44% became single-call recoveries.
• Strong model: parity, with a net 1.4K tokens saved per task.

Teknium later put the weaker-model improvement at “like 25%” in one performance reply. Another reply described NeMo Relay as NVIDIA's version of something like Langfuse for observability, while Teknium's CLI reply clarified that the NVIDIA-linked work still targeted the CLI.

The release list puts “Secrets lockdown with Ironproxy” next to user-facing features, but the docs are more specific. The Ironproxy guide says remote terminal sandboxes hold opaque proxy tokens, while a local host-side iron-proxy daemon swaps those tokens for real credentials at the network boundary.

That boundary has a precise failure mode. The docs say the guarantee depends on the configured trusted proxy boundary, including CA private key and proxy endpoint integrity; stolen CA material or a hijacked endpoint breaks the token guarantee.

The same release also tightened adjacent controls: the v0.20.0 notes mention strict redaction at compaction text boundaries, smart approval allowlist suggestions, consecutive-denial circuit breakers, and a new approval gate for docker/podman daemon-redirect commands.

The desktop docs say Hermes Desktop uses the same agent core, config, API keys, sessions, skills, and memory as the CLI and gateway. Teknium's Intel Mac tip adds the practical installer path: install the CLI, run hermes desktop, and let it compile locally.

The release turns the desktop into a plugin surface. Teknium's release inventory names plugins and Kanban for the desktop GUI, while the plugin-interface tracker frames the longer-term goal as moving long-queued community PRs into stable, publishable plugins.

The CLI got the agent-harness basics engineers keep asking for. Teknium's plan-mode reply says Hermes has /plan <prompt> plus yolo, smart, and always-ask approvals, while the v0.20.0 release notes list ! shell mode, /init, /diff, /context, /focus, prompt stashing, and hermes import-agent for Claude Code or Codex CLI setups.

Update mechanics stayed simple in the thread: Teknium's update reply told one user to run hermes update and type N.

The Agent2Agent piece is now a bundled protocol surface. The Hermes A2A docs say the plugin works both ways: Hermes can call other A2A agents as tools, and other agents can send tasks to Hermes over HTTP.

The same docs describe A2A v1.0 as the open Agent2Agent protocol stewarded by the Linux Foundation, with interoperability across another Hermes instance, LangChain, CrewAI, Google ADK agents, or anything built on the official a2a-sdk.

Outbound webhooks cover the system-integration side. The release notes say Hermes can push signed lifecycle events for session activity, turn completions, and tool events to registered HTTP endpoints.

Teknium said Qwen 3.8 Max and a local 27B Qwen 3.8 are coming, with Hermes Agent featured in Alibaba Qwen's release video. Alibaba_Qwen followed up that it was looking forward to Qwen3.8-Max on Hermes Agent in its reply.

The quickstart puts a hard floor under that local-model story. The Hermes quickstart says local or self-hosted setups should verify endpoint, model name, and context length, and its setup guidance requires at least 64K tokens of context for multi-step tool calling and memory-heavy operation.

Practitioners were already reading the release through the open-frontier lens. Omar said Qwen3.8-Max on Hermes Agent makes him question how much open frontier models have caught up with closed ones in his Qwen reaction. Teknium also tied the efficiency batch to cheaper local/open runs, calling it a fit for DeepSeek V4 Flash 0731 in a follow-up.

ai_for_success's IRIS demo is the clearest example of Hermes as a background runtime instead of a chat window. In the original demo, he says he was talking to IRIS through Bluetooth earbuds while Hermes did the work from another room.

The project is open source, according to ai_for_success's repo reply. In a model-stack reply, he said the setup used Gemini Live for IRIS and Hermes with Gemini 3.6 Flash or Kimi K3.

IRIS also has its own session behavior. ai_for_success's sleep-mode reply says it can sleep after 30 seconds of inactivity, wake on “Hey Iris,” and auto-wake when Hermes completes a task.