Hermes Agent v0.19.0: 80% Faster First Token, Terminal Billing, Password Managers, and GPT-5.6
Time-to-first-token down ~80% across platforms, /subscription and /topup in the terminal, smart approvals on by default, Bitwarden/1Password integration, live subagent transcripts, and a big desktop overhaul.
Two months after v0.14.0 — the last release we covered here — Hermes Agent has raced through five versions and landed on v0.19.0, codename "The Quicksilver Release". The v0.18.0 → v0.19.0 step alone counts 2,245 commits, 1,065 merged PRs, 3,300 issues closed, and contributions from 450+ people.
The codename fits. The headline this time is speed — the first token shows up about 80% sooner — but the quieter story is reliability: responses that can't get lost, approvals that mostly handle themselves, and credentials that never touch the chat window.
~80% Faster to First Token
Time-to-first-token dropped by roughly 80% across every platform. You type, Hermes answers — the dead air before the reply starts is mostly gone.
Around it, a cluster of related work: reasoning models now stream their thinking live by default instead of going silent until the answer is ready, the TUI renders markdown incrementally as it arrives, and the desktop app picked up 20+ dedicated performance PRs.
If v0.14 made Hermes lighter to install, v0.19 makes it feel lighter to talk to.
Billing Without Leaving the Terminal
Two new commands: /subscription manages your plan, /topup adds balance. No dashboard tab, no context switch — check, pay, keep working.
Small feature, but if you've ever had an agent stall mid-task over quota, you know exactly when you'll use it.
Smart Approvals On by Default
Approval prompts are now reviewed by an LLM first, and the mechanism is enabled by default. Routine, low-risk actions get waved through; genuinely risky ones still stop and ask you.
The practical effect: long agent runs interrupt you a lot less, without turning into "auto-approve everything".
Password Manager Integration
Hermes now talks to Bitwarden and 1Password directly. When a task needs a sign-in or a credential, the password manager steps in — you approve in its UI, and the secret stays out of the conversation.
This closes one of the more awkward gaps in agent workflows: needing to hand over a password to a tool whose whole job is remembering everything you say.
Subagents You Can Watch, Work That Can't Get Lost
Two reliability changes that pair well:
Live subagent transcripts. Delegated background tasks now stream their transcripts as they run. You can peek at what a subagent is doing instead of waiting for its final report — and delegation is durable, surviving restarts.
A durable delivery-obligation ledger. Every reply Hermes owes you is tracked in a persistent ledger until it's actually delivered. Crashes, restarts, and disconnects no longer eat responses.
Providers and Models
The provider list keeps widening:
- Fireworks AI and DeepInfra join as first-class providers
- GPT-5.6 variants — Sol, Terra, Luna, plus Pro — fully integrated
- grok-4.5 reaches GA support
- claude-fable-5 and claude-sonnet-5 supported
- LM Studio gains JIT model loading — local models load on demand instead of up front
Whatever you run against, the pattern from earlier releases holds: bring your own provider, Hermes doesn't care which.
Desktop App Overhaul
The desktop app got its deepest rework yet: a new layout-tree architecture designed so features plug into the shell as contributions, a Capabilities page that consolidates Skills, Tools, and MCP in one place, a Hermes Cloud connection mode, and a session/project color system for telling your work apart at a glance. The whole codebase finished its move to TypeScript.
Gateway and Fleet
For people running Hermes as shared infrastructure: per-session turn leases and conversation-scope funneling keep concurrent traffic orderly, each channel can now override the model and system prompt independently, relays support OIDC client-credentials provisioning, and webhooks gained payload filtering plus route scripts.
Multiplex gateways also get profile-based message routing — one gateway, many identities, each conversation routed to the right profile.
Sessions and Export
Sessions can now be exported to Markdown, Quarto, HTML, and trace formats, with redaction options for sensitive data on the way out. Session pruning and bulk archive round out the housekeeping side.
Security Hardening
A steady pass across the credential surface: hardened credential storage, webhook body-size caps, broader token redaction across providers, and sanitized subprocess environments. Nothing flashy — exactly the kind of work you want to see in every release.
How to Upgrade
v0.19.0 is backward compatible. Fresh install:
pip install hermes-agent
Already on Hermes:
hermes update
Full changelog on the GitHub release page.