Hermes Agent Now Starts 80% Faster — Here Is What Changed in v0.19

Hermes Agent v0.19.0 landed on July 20 with over 2,200 commits, 1,065 merged pull requests, and roughly 300,000 lines of new code. The headline change: Hermes now starts about 80% faster than it did in v0.18.0. But speed is just the spine of this release. The rest includes smart approvals that reduce permission fatigue, password manager integration so API keys no longer need plaintext config files, live subagent transcripts, and a delivery ledger that prevents completed responses from being silently lost.

If you use Hermes for anything beyond casual prompts, this is the release to update to.

The speed changes: why Hermes feels faster

Hermes Agent v0.19.0 attacks startup time in two stages.

Stage 1: Cold start dropped from ~4.3 seconds to ~0.9 seconds. Before this release, launching Hermes showed an “Initializing agent…” screen for over four seconds before your first prompt even reached the model. That delay is now under one second, and it applies everywhere — the command-line tool, the gateway server, the TUI, the desktop app, and cron jobs. The improvement comes from moving work off the critical startup path and lazy-loading components that are not needed for the first turn.

Stage 2: Reasoning streams live by default. When Hermes uses a reasoning model — one that “thinks” before answering — the thinking process used to be invisible until the full answer was ready. You would see a spinner for 20-30 seconds with no feedback. Now, Hermes streams the reasoning tokens as they are generated. You see the model’s thinking process appear in real time, which makes long waits feel shorter and lets you catch mistakes early.

Desktop app performance overhaul

The desktop application received over 20 performance-focused pull requests:

14× faster streaming markdown rendering — long responses no longer cause the UI to stutter. The markdown splitter that processes each token was rewritten to avoid redundant re-renders.
Virtualized diffs — reviewing large file changes used to freeze the diff pane. Now only the visible portion of the diff is rendered, keeping the UI responsive even with hundreds of changed lines.
Faster session switching — switching between conversation sessions no longer triggers a full layout recalculation. Profile backends pre-warm on hover intent, and hidden panes mount at idle instead of during cold start.

The net effect: the desktop app feels like a native application under load, even with transcripts that span hundreds of messages and multiple large file changes.

Smart approvals: fewer permission prompts by default

One of the most common complaints about coding agents is approval fatigue. Every time the agent wants to run a command, you get a prompt asking whether to allow it. For a typical coding session, that might mean approving 50-100 commands manually.

Hermes v0.19.0 changes the default behavior: an independent LLM reviewer now evaluates flagged commands automatically instead of asking you to approve each one. The reviewer looks at the specific command in context, decides whether it is safe, and either allows or denies it. Each verdict covers only that exact command — a later command matching the same pattern gets its own review.

How smart approvals work

When Hermes encounters a command it considers risky — a destructive git operation, a file deletion, a network request to an unfamiliar host — the smart approval system sends the command to a lightweight review model. That model evaluates the command against your session context and returns a verdict.

If the command is clearly safe (like git status or ls), it runs without prompting. If it is clearly dangerous (like rm -rf /), it gets denied. The gray area commands — the ones that depend on context — get routed through the reviewer, which considers what you asked for and whether the command makes sense.

You can still manually approve or deny any command. The smart approval system is a default, not a lock. And you can override it for specific command patterns using the new deny rules.

User-defined deny rules

New in v0.19.0: you can set rules that block specific command patterns even under YOLO mode (the mode where everything runs automatically). This gives you guardrails without full approval overhead. For example, you might allow the agent to run any command automatically but deny any command that modifies files outside your project directory.

The /deny <reason> command lets you refuse a command and tell the agent why. Instead of just rejecting it and having the agent guess what went wrong, you explain the problem so it can try a different approach.

Password manager integration: Bitwarden and 1Password

API keys for language models, search engines, and other services traditionally live in plaintext .env files or environment variables. This works for personal development but creates problems for teams: secrets leak into version control, get copied between machines, and are difficult to rotate.

Hermes v0.19.0 adds a SecretSource interface that plugs into Bitwarden and 1Password. When Hermes needs an API key, it fetches it directly from your vault instead of reading it from a file. Multiple vaults can be enabled simultaneously with deterministic precedence — if the same key exists in two vaults, Hermes uses the one you configured first.

The feature supports 1Password’s op:// reference format, so you can reference secrets like op://Personal/OpenAI/api_key in your config. Bitwarden secrets are fetched by name or URI. Both providers work simultaneously, so you can keep some secrets in Bitwarden and others in 1Password.

For beginners: this means you can stop putting API keys in plain text. Install the Bitwarden or 1Password CLI, point Hermes at your vault, and it handles the rest.

Live subagent transcripts

When Hermes delegates a task to a subagent — a child process that handles a specific piece of work — the subagent’s activity is now visible in real time. The delegate_task function returns a transcript file you can follow with tail -f while the subagent runs.

Every tool call, every result, every streamed reply appears in the transcript. This is useful for debugging subagent behavior and for monitoring long-running background tasks.

The transcripts are durable too. If Hermes restarts mid-run, subagent results are restored through an ownership-checked ledger instead of vanishing. This closes a gap where background delegation results could be silently lost if the gateway crashed between dispatch and completion.

Profile-based message routing

A single Hermes gateway can now serve multiple profiles from one bot token. If you use Hermes through Discord, you can route different servers or channels to different profiles — each with its own configuration, skills, memory, and secrets.

This is useful for separating work and personal contexts. Your work server routes to a profile with your work API keys and project-specific skills. Your personal server routes to a different profile with personal credentials. One bot token, two completely isolated setups.

New models and providers

The release adds support for several new models and providers:

Fireworks AI as a first-class provider with cost estimation
DeepInfra as a first-class provider
GPT-5.6 (all variants: Sol, Terra, Luna, Pro) wired end-to-end
Grok 4.5 (general availability)
Moonshot Kimi K3
Claude Fable 5 and Claude Sonnet 5
Tencent Hy3 (general availability)
LM Studio JIT model loading for local setups

Reasoning effort also gained max and ultra levels for models that support them, with per-model and per-task overrides configurable from the CLI or desktop app.

Session export and data portability

hermes sessions export now writes conversation history in Markdown, Quarto, HTML, prompt-only, and Hugging Face-ready trace formats. You can filter by age, workspace, or platform, and opt into a --redact pass that scrubs secrets from the output.

Your conversation history becomes a real dataset you can analyze, share, or use for fine-tuning — not a black box locked inside Hermes’s internal storage.

How to update

If you installed Hermes via npm:

npm install -g hermes-agent@latest

Verify with hermes --version — you should see v0.19.0 or later. The speed improvements, smart approvals, and new providers work out of the box. For password manager integration, install the Bitwarden or 1Password CLI and configure the secretSource in your Hermes settings.

The bigger picture

Hermes v0.19.0 is a release that attacks the friction points experienced users accumulate over months of daily use. The speed improvements make it feel responsive from the first token. Smart approvals reduce the constant yes/no prompts without removing control. Password manager integration makes secrets management practical instead of theoretical. And the durable delivery ledger means completed work is not silently lost.

For developers choosing between Hermes, Claude Code, Cline, and the other coding agents on the market, this release narrows the gap on startup speed and widens it on features like multi-profile routing and secret management. The comparison is worth making on your own workflow, not on paper.