Hermes v2026.3.17 Update: Hermes Agent v0.3.0 (v2026.3.17)

This release turns Hermes from a clever assistant into a real-time teammate—with plugins, voice, browser control, and a smarter approval system.

What’s New

🎉 Streaming: Watch it think in real time

Previously, Hermes would wait until it finished a whole response before showing you anything. Now, with unified streaming, you see tokens appear as they’re generated—like watching a friend type a message live instead of waiting for the whole paragraph. This works across the CLI and all gateway platforms. (#1538)

🧩 Plugins: Drop in, plug in

You no longer need to fork the project to add custom tools. Just drop a Python file into ~/.hermes/plugins/ and Hermes picks it up automatically. Think of it like adding apps to your phone—no rooting required. (#1544, #1555)

🏦 Native Anthropic Provider

Hermes can now talk directly to Anthropic’s API—no middleman. It even auto-discovers Claude Code credentials and supports OAuth PKCE flows. If you use Claude, this is like getting a direct line instead of going through a switchboard. (#1097)

🛑 Smart Approvals + `/stop`

Inspired by Codex, Hermes now learns which commands are safe and remembers your preferences. You’ll get fewer “are you sure?” prompts over time. And if you need to bail out instantly, /stop kills the current run immediately. (#1543)

🧠 Honcho Memory Integration

Hermes now writes memories asynchronously (no lag), supports configurable recall modes, and keeps sessions separate per user in gateway mode. It’s like giving Hermes a notebook that it writes in quietly, in the background. (#736)

🎤 Voice Mode

Push-to-talk in the CLI, voice notes in Telegram/Discord, and even Discord voice channel support. Local transcription uses faster-whisper, so your audio stays private. (#1299, #1185, #1429)

⚡ Concurrent Tool Execution

Multiple independent tool calls now run in parallel via ThreadPoolExecutor. If Hermes needs to check three websites at once, it does—cutting latency dramatically. (#1152)

🕵️ PII Redaction

Enable privacy.redact_pii and Hermes automatically scrubs personal info (emails, phone numbers, etc.) before sending context to LLM providers. Your data stays yours. (#1542)

🌐 `/browser connect` via CDP

Attach Hermes to a live Chrome instance through Chrome DevTools Protocol. Debug, inspect, and interact with pages you already have open—like handing Hermes your browser tab, not a screenshot. (#1549)

☁️ Vercel AI Gateway Provider

Route Hermes through Vercel’s AI Gateway for access to their model catalog and infrastructure. (#1628)

🧭 Centralized Provider Router

The entire provider system was rebuilt around a single router, making it easier to add new providers and switch between them.

Improvements

Persistent shell — Hermes now keeps a session alive across commands, so cd and export actually stick.
ACP IDE integration — Works with coding agents in your IDE.
Better Telegram/Discord handling — Smoother voice notes and channel support.

Fixes

50+ bug fixes across every platform. Notable ones include:

Fixed streaming glitches on Windows terminals
Resolved memory race conditions in multi-user gateway mode
Corrected plugin loading order so your custom tools override defaults predictably

What This Means for You

This release makes Hermes feel faster, more personal, and more private. Streaming makes it feel alive. Plugins make it yours. Voice makes it hands-free. And the approval system means fewer interruptions as it learns your workflow.

If you’ve been waiting for a reason to switch from a static assistant to a true agent, this is it.

📎 Official release notes: https://github.com/NousResearch/hermes-agent/releases