Last updated on

Hermes v0.19.0 Quicksilver Release: 80% Faster First Token, 11 Standout Features

Hermes Agent v0.19.0, codenamed Quicksilver, shipped on July 20, 2026. It is one of Nous Research’s largest releases to date: since v0.18.0, the community landed roughly 2,245 commits, 1,065 merged PRs, 3,300 closed issues, and contributions from 450+ people. Our site’s v0.19.0 release notes put it plainly: “Hermes is the messenger god, and this window we made him move like it.”

This post is not a copy-paste of the changelog. I will break the update into the 11 features ordinary developers should care about, why they matter, how to use them, and what practical impact you can expect. For the full technical changelog, see our v0.19.0 release notes.

1. First-token latency drops ~80% everywhere

The headline number: time-to-first-token (TTFT) fell from ~4.3s to ~0.9s, an ~80% cut. And it applies across the CLI, gateway, TUI, desktop, and cron.

How did that happen? The team removed several bottlenecks from the cold-start path:

  • Discord capability detection moved off the critical path via a 24-hour disk cache plus background refresh
  • Ollama probe is skipped for known non-Ollama providers
  • Blocking agent-init work was removed from cold start
  • Prompt-build caching and mtime-cached timezone resolution were added

This is not one big optimization; it is dozens of small cuts that together remove the “Initializing agent…” delay. If Hermes ever felt like it took a deep breath before answering, that breath is gone.

2. Reasoning models stream their thinking live

display.show_reasoning is now ON by default. With reasoning models such as Claude 3.7 Sonnet or the o3/o4 series, you no longer stare at a spinner for 30 seconds. You see the model think step by step. The response box also paints per token instead of per line, which makes the whole interaction feel more responsive.

3. Desktop app gets a 20-PR performance overhaul

Hermes Desktop underwent a focused speed-and-smoothness sprint:

  • Streaming markdown rendering uses 14× less CPU via incremental block lexing
  • Giant diffs are virtualized, so the review pane no longer freezes on full Shiki highlighting
  • Session switching is snappy even with huge transcripts; the layout-thrash cascade was eliminated
  • Startup serialization and per-turn REST amplification were reduced
  • Profile backends pre-warm on hover intent; boot-hidden panes mount at idle
  • Sidebar and tool rows no longer re-render on every token during streaming

The result is a desktop experience that feels native under load, even with busy agents and large conversations.

4. Manage your Nous plan from the terminal: /subscription and /topup

Changing subscriptions used to require a trip to the billing website. Now it lives in the TUI or classic CLI:

/subscription
/topup

/subscription shows your current plan, remaining allowance, exact upgrade cost (e.g., “Pay $46.30 & upgrade now”), downgrade effective date, and lets you apply or undo changes. A matching billing tab was added to the desktop app. For power users, this is a significant quality-of-life win.

5. Smart approvals: let an LLM judge flagged commands

Hermes already asked for approval before running flagged commands. In v0.19.0, smart approvals are the default. An independent LLM reviewer classifies the command as safe, dangerous, or uncertain. Safe commands are auto-approved; dangerous ones are denied; uncertain ones are escalated to you. Each command is reviewed individually, so a later match does not get a free pass.

Combined with user-defined deny rules that block commands even in yolo mode and /deny <reason> which tells the model why you refused, this reduces approval fatigue without giving up control.

approvals:
  deny:
    - "rm -rf *"
    - "DROP DATABASE *"
    - "DROP TABLE *"

6. Plug your password manager into Hermes: Bitwarden & 1Password

API keys no longer have to live in a plaintext .env file. A new pluggable SecretSource interface lets Hermes fetch secrets from Bitwarden and 1Password (op:// references) at load time.

secret_sources:
  - provider: bitwarden
    vault: "dev-creds"
    priority: 1
  - provider: onepassword
    vault: "Production"
    priority: 2

Multiple vaults can be active simultaneously with deterministic precedence, conflict warnings, and per-variable provenance. Future vault providers only need to implement the same interface.

7. Watch your subagents work live, with durable background delegation

delegate_task dispatches now return live transcript files you can tail -f from the moment the subagents launch. Every tool call, result, and streamed reply is visible per child. More importantly, background delegation completions are now durable: if the main process restarts mid-run, results are restored and delivered through an ownership-checked ledger instead of vanishing.

This makes it practical to fan out a fleet of subagents and check on them whenever you want, rather than babysitting the terminal.

8. Final answers can no longer be lost to a gateway crash

This is a strong reliability upgrade. If the gateway died between generating your response and confirming delivery to the platform, the answer used to be silently lost — even though you had already paid for the tokens. Final responses are now recorded in a durable delivery-obligation ledger in state.db and redelivered on the next boot.

For Telegram, Discord, Slack, and other gateway channels, this moves delivery from best-effort to at-least-once success.

9. One gateway, many profiles: profile-based routing

A single gateway sharing one bot token can now route specific guilds, channels, or threads to different profiles. Each profile has fully isolated config, skills, memory, and secrets.

For example:

  • Your work Discord server → work profile
  • Your personal Discord server → personal profile
  • All through the same bot token

A second multiplex hardening wave means one misconfigured profile can no longer take down the whole gateway.

10. New providers and the latest frontier models

Model support added in v0.19.0 includes:

  • GPT-5.6 family (Sol / Terra / Luna and Pro variants)
  • grok-4.5 (GA)
  • moonshotai/kimi-k3 (1M context on Kimi Coding endpoints)
  • claude-fable-5 / claude-sonnet-5
  • tencent/hy3 GA
  • GLM-5.2, Upstage Solar, and more

Provider additions include:

  • Fireworks AI as a first-class provider with cost estimation and a #2 slot in the model picker
  • DeepInfra hardened integration
  • LM Studio JIT load mode
  • Bedrock catalog improvements: live context-window probing, 1M context rows, geo-prefix parity

If you frequently switch between cutting-edge models, this release catches you up.

11. Reasoning effort is now a dial: max, ultra, and per-model overrides

reasoning_effort gains max and ultra levels, selectable across every surface and route. You can also:

  • Pin per-model reasoning-effort overrides in config
  • Set per-slot effort in MoA presets
  • Set per-task effort for auxiliary models

Thinking depth becomes a knob you can turn, not a global on/off switch.

Other notable improvements

  • Session export: hermes sessions export writes Markdown, Quarto, HTML, prompt-only, and Hugging Face-ready trace formats, with filters for age/workspace and an optional --redact pass
  • Desktop Capabilities page: Skills, Tools, MCP, and Hub in one place, with test/toggle and log filtering
  • Web Dashboard: safe session import, WhatsApp pairing, Discord toolset management, terminal keep-alive, and more
  • Security hardening: Vertex credential scoping, six P1 browser/MEDIA/.env hardening PRs, webhook body-size caps, CI untrusted-ref protection
  • CLI utilities: /model --once for one-turn model override, stacked skill invocations (/skill-a /skill-b), and --safe-mode troubleshooting

How to upgrade

If you already have Hermes installed:

hermes update
# or
npm update -g @nousresearch/hermes-agent

After upgrading, run a quick health check:

hermes config doctor
hermes plugins list
hermes mcp list

Summary

The defining change in Hermes v0.19.0 Quicksilver is that it got faster and stayed reliable. An ~80% first-token improvement, 14× faster desktop streaming markdown, and crash-proof gateway delivery are not flashy features; they are the foundation that decides whether an autonomous AI agent becomes part of your daily workflow.

On top of that foundation, v0.19.0 adds enterprise-grade capabilities: terminal billing, password-manager integration, smart approvals, subagent monitoring, a delivery ledger, and multi-profile routing. If you already run Hermes in production or as a team gateway, this release is a strong upgrade.

If you have not tried it yet, start with /subscription and a hermes update, then let the new speed handle your next task.