Hermes Agent changelog
v0.20.0 August 3, 2026
The "Herald" release — Hermes gets a voice you can interrupt, talks to other agents over A2A v1.0, and cites its sources. Rolls up everything since v0.19.0, including the v0.19.1 patch window.
Added
- Streaming conversational voice: Hermes speaks clause-by-clause as the reply generates, and you can cut in mid-sentence — it stops, listens, and the model is told you interrupted.
- Open-vocabulary wake words detected on-device, so no audio leaves your machine while it waits. Different wake phrases can route to different profiles.
- Voice notes in and TTS replies out across WhatsApp, Feishu, DingTalk, LINE, QQ, Photon and Weixin. Speech-to-text is now its own configurable `hermes tools` category.
- Agent-to-agent messaging via A2A v1.0.
- Signed outbound webhooks, so Hermes can announce events to your own systems.
- Grounded research with verifiable citations and fact-checking.
- Desktop app becomes a platform: artifacts with live preview, a plugin SDK, quick-entry from anywhere, and multiple windows.
- New CLI commands: `!` shell mode, `/init`, `/diff`, `/context` and `/focus`.
Fixed
- Tools now recover from their own failures instead of handing the model a dead end to guess around.
- Gentler, smarter context compression.
- Absorbs the entire v0.19.1 salvage window across gateway, voice, desktop and installer.
v0.19.1 July 30, 2026
Patch release with no headline features — it exists to give Docker images, hosted deployments and fresh installs a stable tag covering the ~1,000 PRs merged since v0.19.0.
Added
- Buzz/Nostr channel support.
- FLUX3 video generation and delivery.
Fixed
- Bug-fix waves across the gateway, the voice subsystem, the desktop app and the installer.
- Telegram media reliability.
- Voice-mode regressions introduced during the v0.19.0 window.
v0.19.0 July 20, 2026
The "Quicksilver" release — first-turn time-to-first-token drops roughly 80% across every platform, plus password-manager secrets and smart approvals on by default.
Breaking
- Smart approvals are now on by default. Automation that previously ran unattended may start waiting for a confirmation it never used to need — review your approval settings before assuming the agent has hung.
Added
- Live reasoning streams, so you can watch the agent think instead of staring at a blank prompt.
- Secrets can be pulled from Bitwarden and 1Password instead of sitting in plaintext env files.
- Subscription management from the terminal.
- Durable response delivery — replies survive a transient gateway disconnect rather than vanishing.
Fixed
- Substantial desktop and TUI responsiveness work across the board.
v0.18.2 July 8, 2026
Same-day patch on v0.18.1, picking up the WhatsApp Baileys dependency fix that tagged Docker builds needed.
Fixed
- fix(whatsapp): unpin Baileys from git commit, use published 7.0.0-rc13
v0.18.1 July 8, 2026
A patch tag cut so downstream consumers — Docker images, hosted deployments, fresh installs — had something stable on top of the v0.18.0 window.
v0.18.0 July 1, 2026
The "Judgment" release — the team spent a week and a half closing every P0 and P1 in the repo. Roughly 700 highest-priority items, all of them shut.
Added
- Mixture-of-Agents is now a first-class model you can pick
- See every model's reasoning, then watch the answer stream in
- The agent verifies its own work — "done" means proven, not claimed
- /learn — turn anything into a reusable skill by describing it
- /journey — a playable timeline of everything Hermes has learned about you
- Delegate a pile of work and keep going — background fan-out
- First-class coding Projects in the desktop app
- Run Hermes at scale — scale-to-zero and drain coordination
- Cheaper self-improvement — smarter background review
- Compose your next prompt in your editor — /prompt
- Google Vertex AI — Gemini through your GCP service account, no static key
- Presets as selectable virtual models
- /moa is now one-shot sugar
Fixed
- Security round
v0.17.0 June 19, 2026
The "Reach" release — v0.16.0 put Hermes on your desktop; this one is about how far that reach extends. iMessage, the Raft agent network, background subagents and team administration.
Added
- Hermes reaches iMessage — Photon Spectrum, no Mac relay required
- Raft — Hermes joins the Raft agent network as a gateway channel
- A substantially more capable desktop app
- Background / async subagents — delegate work and keep going
- Edit images, not just generate them — image-to-image in image_generate
- Automation Blueprints — schedule things without learning cron
- Cursor's Composer model, through your xAI Grok subscription
- Full profile builder in the dashboard
- Skills Hub browser rehaul
- The memory tool got a major upgrade — atomic batch operations
- Secure dashboard login
- Official WhatsApp Business Cloud API adapter
- Rich text for Telegram — Bot API 10.1 rich messages
- Curator cost optimization — no aux-model spend on routine runs
v0.16.0 June 6, 2026
The "Surface" release — a real native desktop app for macOS, Linux and Windows, built across 100 PRs in a single week. Not a terminal wrapper.
Added
- Hermes Desktop — a real native app, not a terminal wrapper
- Run the desktop app against a remote Hermes — sign in with OAuth or username/password
- The web dashboard is now a full admin panel — configure everything from the browser
- Hermes Desktop speaks Simplified Chinese — full 简体中文 in the chat GUI
- Leaner default skill set — Hermes ships only what you actually need
- NVIDIA/skills is now a built-in trusted skills tap
- Quick Setup via Nous Portal — from install to first message in seconds
- Fuzzy model picker, everywhere — type a few letters, find your model
- /undo [N] — take back the last N turns
- Choose your default interface — cli or tui
- Simplified Chinese (简体中文) translation
Fixed
- CVE-2026-48710 (Starlette BadHost)
v0.15.1 May 29, 2026
The "Patch" release — a same-day hotfix for v0.15.0, headlined by the dashboard infinite-reload loop that hit anyone running in loopback mode: Docker, hosted Hermes and fresh installs.
Fixed
- Dashboard 401 reload loop fixed
- Docker dashboard --insecure is now an explicit env opt-in, never derived from bind host
- MCP bare command resolution under Docker
- Kanban worker can be killed again
- Full skills.sh catalog (858 → 19,932 entries)
- /api/auth/me 401 no longer triggers reload loop
- Skills page source pills + category sidebar restored
- insecure is now explicit opt-in via env var
- Dashboard test suite repaired
- arm64 PR builds skip the GHA cache
- Bare npx/npm/node resolve against /usr/local/bin
- Worker SIGTERM actually terminates the process
- Workers receive images referenced in task bodies
v0.15.0 May 28, 2026
The "Velocity" release — the 16,083-line run_agent.py collapses to 3,821 across 14 modules, and Kanban grows into a real multi-agent platform over 104 PRs.
Added
- The Big Refactor — run_agent.py is no longer 16,000 lines
- Kanban grew into a real multi-agent platform — 104 PRs end to end
- Cold-start perf wave keeps going — another second saved, 47% fewer per-turn function calls
- session_search rebuilt — no LLM, no cost, 4,500× faster
- Promptware defense — Brainworm-class attacks blocked at three chokepoints
- Bitwarden Secrets Manager — one bootstrap token replaces every per-provider API key
- ntfy as the 23rd messaging platform — push notifications without an account
- Skill bundles — /<name> loads multiple skills at once
- TUI session orchestrator — multiple live sessions in one TUI window
- Two new image_gen providers — Krea 2 Medium + Large, FAL ported to plugin
- Nous-approved MCP catalog with interactive picker
- OpenHands orchestration skill
Fixed
- Deep xAI integration round — Web Search plugin, OAuth proxy upstream, May 15 retirement detection, natural TTS, security hardening
v0.14.0 May 16, 2026
Hermes becomes pip-installable, so you no longer have to pipe a shell script into your machine to try it.
Added
- `pip install hermes-agent` as a first-class install path alongside the curl installer.
Fixed
- Assorted installer reliability fixes around dependency detection.
v0.13.0 May 7, 2026
Multi-agent Kanban arrives, the agent stops forgetting what you asked it to do, and it learns to watch video and clone a voice.
Added
- Multi-agent Kanban — delegate to an AI team that actually finishes
- /goal — the agent doesn't forget what you asked it to do
- Show it a video
- Clone a voice
- Hermes speaks your language
- Google Chat — the 20th messaging platform
- Sessions survive restarts
- Checkpoints v2
- The agent lints its own writes
- no_agent cron mode — script-only watchdog
- Platform allowlists everywhere
- Providers are now plugins
- API server — long-term memory per session
Fixed
- Security wave — 8 P0 closures
v0.12.0 April 30, 2026
The Autonomous Curator lands and the self-improvement loop gets a substantial upgrade, alongside a wave of new inference providers and messaging platforms.
Added
- Autonomous Curator
- Self-improvement loop — substantially upgraded
- Skill integrations — major expansion
- LM Studio — first-class provider
- Four more new inference providers
- Pluggable gateway platforms + Microsoft Teams
- Tencent 元宝 (Yuanbao) — 18th messaging platform
- Spotify — native tools + bundled skill + wizard
- Google Meet plugin
- hermes -z one-shot mode + hermes update --check
- Models dashboard tab + in-browser model config
- Remote model catalog manifest
- Native multimodal image routing
- Gateway media parity
v0.11.0 April 23, 2026
A rebuilt Ink-based TUI, native AWS Bedrock, five new inference paths, and /steer for nudging the agent mid-run.
Added
- New Ink-based TUI
- Transport ABC + Native AWS Bedrock
- Five new inference paths
- GPT-5.5 over Codex OAuth
- QQBot — 17th supported platform
- Plugin surface expanded
- /steer — mid-run agent nudges
- Shell hooks
- Webhook direct-delivery mode
- Smarter delegation
- Auxiliary models — configurable UI + main-model-first
- Dashboard plugin system + live theme switching
- Dashboard polish
v0.9.0 April 13, 2026
The local web dashboard arrives, along with Fast Mode, Android support via Termux, and a comprehensive security hardening pass.
Added
- Local Web Dashboard
- Fast Mode (/fast)
- iMessage via BlueBubbles
- WeChat (Weixin) & WeCom Callback Mode
- Termux / Android Support
- Background Process Monitoring (watch_patterns)
- Native xAI & Xiaomi MiMo Providers
- Pluggable Context Engine
- Unified Proxy Support
- hermes backup & hermes import
- 16 Supported Platforms
- /debug & hermes debug share
- Native xAI (Grok) provider
Fixed
- Comprehensive Security Hardening
v0.8.0 April 8, 2026
Background processes learn to notify you when they finish, models become switchable live with /model, and Gemini gets a native provider.
Added
- Background Process Auto-Notifications (notify_on_complete)
- Free Xiaomi MiMo v2 Pro on Nous Portal
- Live Model Switching (/model Command)
- Self-Optimized GPT/Codex Tool-Use Guidance
- Google AI Studio (Gemini) Native Provider
- Inactivity-Based Agent Timeouts
- Approval Buttons on Slack & Telegram
- MCP OAuth 2.1 PKCE + OSV Malware Scanning
- Centralized Logging & Config Validation
- Plugin System Expansion
- Matrix Tier 1 & Platform Hardening
- Native Google AI Studio (Gemini) provider
- /model command — full provider+model system overhaul
Fixed
- Security Hardening Pass
v0.7.0 April 3, 2026
Memory becomes a pluggable interface, providers gain credential pools, and the agent starts actively blocking secret exfiltration.
Added
- Pluggable Memory Provider Interface
- Same-Provider Credential Pools
- Camofox Anti-Detection Browser Backend
- Inline Diff Previews
- API Server Session Continuity & Tool Streaming
- ACP: Client-Provided MCP Servers
- Gateway Hardening
- Security: Secret Exfiltration Blocking
- Credential pool preserved through smart routing
- Per-turn primary runtime restoration
- developer role for GPT-5 and Codex models
- Google model operational guidance
- Anthropic long-context tier 429 handling
v0.6.0 March 30, 2026
Profiles let you run multiple independent Hermes instances, plus MCP server mode, an official Docker container, and an ordered fallback provider chain.
Added
- Profiles — Multi-Instance Hermes
- MCP Server Mode
- Docker Container
- Ordered Fallback Provider Chain
- Feishu/Lark Platform Support
- WeCom (Enterprise WeChat) Platform Support
- Slack Multi-Workspace OAuth
- Telegram Webhook Mode & Group Controls
- Exa Search Backend
- Skills & Credentials on Remote Backends
- Fix api_mode on provider switch
- Stop silent OpenRouter fallback
- Gemini 3.1 preview models
v0.5.0 March 28, 2026
Nous Portal opens up to 400+ models and Hugging Face becomes a first-class inference provider, with supply-chain hardening alongside.
Added
- Nous Portal now supports 400+ models
- Hugging Face as a first-class inference provider
- Telegram Private Chat Topics
- Native Modal SDK backend
- Plugin lifecycle hooks activated
- Improved OpenAI Model Reliability
- Nix flake
- Supply chain hardening
- Anthropic output limits fix
- Surface lifecycle events
- Thinking-budget exhaustion detection
- Persist reasoning across gateway session turns
v0.4.0 March 24, 2026
An OpenAI-compatible API server, six new messaging platforms, @ context references, and streaming on by default.
Added
- OpenAI-compatible API server
- 6 new messaging platform adapters
- 4 new inference providers
- MCP server management CLI
- Gateway prompt caching
- Context compression overhaul
- Streaming enabled by default
- /statusbar
- /queue
- /permission
- /browser
- /cost
v0.3.0 March 17, 2026
Unified streaming infrastructure, a first-class plugin architecture, a native Anthropic provider, and the first voice mode.
Added
- Unified Streaming Infrastructure
- First-Class Plugin Architecture
- Native Anthropic Provider
- Smart Approvals + /stop Command
- Honcho Memory Integration
- Voice Mode
- Concurrent Tool Execution
- PII Redaction
- /browser connect via CDP
- Vercel AI Gateway Provider
- Centralized Provider Router
- ACP Server (IDE Integration)
- Persistent Shell Mode
- Agentic On-Policy Distillation (OPD)
v0.2.0 March 12, 2026
The multi-platform messaging gateway, the MCP client and the skills ecosystem all land — the release where Hermes becomes reachable from everywhere.
Added
- Multi-Platform Messaging Gateway
- MCP (Model Context Protocol) Client
- Skills Ecosystem
- Centralized Provider Router
- ACP Server
- CLI Skin/Theme Engine
- Git Worktree Isolation
- Filesystem Checkpoints & Rollback
- 3,289 Tests
- ASCII Art
- ASCII Video
- DuckDuckGo Search
- Solana Blockchain
- AgentMail