Hermes Agent v0.19.0 shipped on July 20, 2026 with over 2,200 commits and 3,300 closed issues since the last major release. The headline: first-token startup time dropped roughly 80 percent across CLI, gateway, TUI, desktop, and cron. But speed is only the spine — the release also adds Bitwarden and 1Password secret sources, smart approval defaults, live subagent transcripts, a delivery-obligation ledger that prevents lost responses, and a desktop app that feels like a native application under load.
If you have been waiting for Hermes to feel fast and safe enough for daily use, this is the release that gets it there.
Why the speed jump matters
Before v0.19.0, cold-starting Hermes took about 4.3 seconds before your first message even reached the model. Discord capability detection, Ollama probes, and agent-init blocking all sat on the critical path. Now those are cached, deferred, or removed entirely. Cold submit-to-dispatch sits at roughly 0.9 seconds — an 80 percent cut that you feel on every platform.
The second speed win is perceptual. Reasoning models now stream their thinking live by default instead of showing a spinner for 30 seconds. The response box paints per token instead of per line, and prompt-build caching means repeated turns hit a warm cache. For beginners, the practical effect is simple: Hermes no longer takes a deep breath before answering.
Password manager integration: no more plaintext API keys
If you have been storing API keys in a plaintext .env file, you are not alone — but it is a security risk. Hermes v0.19.0 introduces a pluggable SecretSource interface that fetches secrets from Bitwarden and 1Password at load time.
Here is what that looks like in practice:
- You reference a Bitwarden or 1Password item in your Hermes config (using
op://syntax for 1Password, for example). - At startup, Hermes fetches the actual secret values from your vault.
- Multiple vaults can be enabled simultaneously with deterministic precedence and per-variable provenance tracking.
- Future vault providers drop in as plugins, so if you use a different password manager, support may come later.
The result: API keys stay encrypted in your vault instead of sitting in a file that any tool on your machine can read. This alone is worth the update if you care about security.
Smart approvals reduce the permission pop-ups
By default, Hermes now runs an independent LLM reviewer when it wants to execute a flagged command. Instead of asking you to approve every single dangerous-looking action, the reviewer assesses each command individually and decides whether it is safe enough to run.
This works alongside two new features:
- User-defined deny rules — You can set rules that block certain commands even when Hermes is in “yolo” (auto-approve) mode. Think of these as hard stops that cannot be overridden.
/deny <reason>— When you refuse a command, you can tell Hermes why. The agent uses that feedback to adjust its approach instead of retrying the same thing.
For beginners, the net effect is fewer interruptions. You approve the genuinely risky stuff and let the tool handle routine operations on its own.
Desktop app: 14x faster markdown rendering
The desktop app received over 20 performance pull requests in this release. The biggest wins:
- Incremental block lexing for streaming markdown cuts the markdown splitter’s CPU usage by 14x on long replies.
- Virtualized diffs in the review pane mean giant code diffs no longer freeze the screen.
- Session switching no longer thrashes the layout, even with huge transcripts.
- Sidebar and tool-row rendering is batched instead of re-rendered per token during streaming.
If you have been using the Hermes desktop app and it felt sluggish with long conversations or busy agents, this update fixes that. Under load, it now feels like a native application.
Subscription management from the terminal
Changing your Nous subscription used to require opening a browser. Now you can run /subscription directly in the TUI or CLI to:
- See your current plan and remaining usage allowance.
- Preview upgrade costs (“Pay $46.30 and upgrade now”) or schedule downgrades.
- Apply changes immediately with scheduled-change banners and undo support.
The desktop app has a matching billing settings tab. Your wallet stays at the keyboard.
Live subagent transcripts
When Hermes delegates a task to a subagent, the dispatch now returns a live transcript file you can watch with tail -f. Every tool call, result, and streamed reply logs to one human-readable file per child agent.
Background delegation completions are also durable now. If the process restarts mid-run, results are restored through an ownership-checked ledger instead of vanishing. You can fan out a fleet of subagents, watch any worker live, and never lose results to a crash.
Delivery ledger: responses survive gateway crashes
Before this release, if the gateway died between generating a response and confirming delivery, that answer was silently lost. You paid for the turn but never got the output. This affected Telegram, Discord, Slack, and every other channel.
v0.19.0 records final responses in a durable ledger in state.db around the platform send and redelivers them on the next boot. A finished answer can no longer be lost.
Profile-based routing: one gateway, many configs
A single multiplexed gateway sharing one bot token can now route specific servers, channels, or threads to different Hermes profiles. Each profile has fully isolated config, skills, memory, and secrets.
For example, point your work Discord server at a work profile and your personal server at a personal profile — from one bot instance. One misconfigured profile can no longer bring down the whole gateway.
New providers and frontier models
The release adds Fireworks AI and DeepInfra as first-class providers, plus Upstage Solar. Model catalogs now include GPT-5.6 variants (Sol, Terra, Luna, and Pro), grok-4.5, kimi-k3, claude-fable-5, claude-sonnet-5, and tencent/hy3. LM Studio gets JIT model loading for local setups.
Reasoning effort gained max and ultra levels with per-model and per-slot control, so you can set your advisors to think hard while keeping your synthesizer fast.
Session export for data portability
hermes sessions export now writes Markdown, HTML, prompt-only, and Hugging Face-ready trace formats. You can filter by age, workspace, and platform, opt in to secret-scrubbing with --redact, and stitch compacted-session lineage into one logical export. Your conversation history becomes a real dataset.
How to update
hermes update # existing installs
pip install -U hermes-agentThe bottom line
Hermes Agent v0.19.0 is the release where speed, security, and reliability stop being aspirational and start being defaults. The 80 percent faster startup means less waiting. Password manager integration means API keys stay encrypted. Smart approvals mean fewer interruptions. The delivery ledger means responses do not disappear. And the desktop app finally keeps up with long sessions. If you have been on the fence about Hermes as a daily driver, this is the version to try.