Hermes Agent v0.20.0: It Finally Learned to Talk Back

Hermes Agent v0.20.0: It Finally Learned to Talk Back

Let me tell you how I stumbled onto this release.

I use Hermes every day for work. Been using it since last year. There is one thing that has been driving me crazy the entire time: the voice mode. Old Hermes would get your voice message, wait for the whole response to generate, then play you a complete audio file. You could not interrupt it. It would not think out loud. Using it felt like leaving a voicemail, not having a conversation.

v0.20.0 fixed that for good.

The release is codenamed Herald. In Greek mythology, Hermes was the messenger of the gods . he delivered words, made announcements, and guided travelers. The name fits: this version of Hermes learned to talk (real-time voice), learned to communicate with other agents (A2A protocol), learned to push events outward on its own (signed webhooks), and learned to cite verifiable sources for its research conclusions (grounded citations).

The numbers tell you the scope: 3,650 commits, 1,400 merged PRs, 1,200 closed issues, 647 contributors. This is not a small window.

Voice: From Voicemail to Real Conversation

This is the change you will notice first.

The old voice mode worked like a voicemail. You say something, Hermes generates the full response, then plays you the audio. No interrupting, no thinking out loud. You just sat there waiting.

v0.20.0 rewrote the model from the ground up. Streaming TTS with interruption. Hermes now synthesizes speech sentence by sentence, speaking while it generates. You can cut in at any point . just start talking and it will stop, listen to you, and the model itself knows it was interrupted. The next sentence starts rendering while the current one is still playing, keeping the wait feel to an absolute minimum.

Coverage: CLI voice mode, desktop app, all gateway platforms that support audio.

Two details worth highlighting.

On-device wake word. You can set a custom wake phrase like “hey Hermes.” Detection runs locally, so audio does not leave your machine while on standby. Different wake words can route to different profiles . say one word to wake your work assistant, another to trigger your home smart device control. Saying “stop” ends the voice conversation from any platform, no keyboard required.

Cross-platform voice. Send a voice message on WhatsApp, Feishu, DingTalk, LINE, QQ, Photon, or WeChat, and Hermes transcribes and replies automatically. All TTS providers share a unified text preprocessing pipeline that strips markdown, code blocks, and URLs from the speech. You will not hear Hermes reading out “hash hash heading” or “backtick print parenthesis.”

Three Tools for Agent Communication

Voice handles communication between humans and Hermes. These three handle communication between Hermes and the outside world.

A2A v1.0. A new bundled plugin implements the Agent-to-Agent protocol. Hermes can discover, talk to, and be driven by other A2A-compatible agents. This closed GitHub issue #514 . one of the oldest feature requests in the repo. If you are building multi-agent systems, Hermes now has a standard protocol to plug into.

Outbound webhooks. Integrating Hermes used to mean polling or monitoring a platform. Now Hermes actively pushes signed lifecycle events to your registered HTTP endpoints: session activity, turn completion, tool events . all with HMAC signature verification so your receiver can confirm authenticity. Hooking up CI, smart home, or a dashboard no longer requires polling. Hermes went from “you ask, it answers” to an event source that can trigger downstream flows.

Grounded citations. The new grounded-citations skill makes Hermes cite verifiable sources for every claim when doing research. These are not model hallucinations . it matches against actual page text and links to the exact evidence location. There is also a fact-checking mode: give it any document or claim and it will tell you which parts checked out, which did not, and which could not be verified. When using Hermes for research, this is the difference between “sounds right” and “I can actually back this up.”

Desktop App: From Chat Client to Platform

v0.19.1 already showed signs of the desktop going IDE-like (command palette, tab management, message replies). v0.20.0 pushes it to the edge of being a full platform.

Artifacts. Versioned cards with a sandboxed live preview pane. HTML or apps Hermes generates can run right in a safe sandbox next to the chat, no copy-pasting to a browser to test. Artifacts are versioned, so you can roll back to a previous version.

Plugin SDK. A real plugin SDK shipped. Kanban (task board system) is the founding plugin. Plugins can use ctx.download to send files to users, float panels, and open multiple GUI windows. The Widget app model also arrived: apps abstracted as state + reducer + render, with three reference implementations and a grid layout engine. Third parties can now build real applications inside the Hermes desktop app, not just chat plugins.

Quick-entry and remote backend. Global hotkey to capture a thought from anywhere in the OS and drop it into any session. Cmd+O opens a folder as a project. The desktop also supports SSH remote backend mode . connect from your laptop to a Hermes instance running on a server, with event-driven real-time sync replacing constant polling.

Desktop UX wave. The editor got an attachment picker (files, folders, links), a composer undo stack, double-ESC to discard drafts, double-Enter to send queued messages. Two-key model switching (Cmd+Shift+M). iMessage-style emoji reactions arrived (opt-in, bidirectional). Session sidebar gained date dividers, a pinned zone, and optional auto-archiving of expired sessions. The 60fps performance push is into its second wave: streaming render cost no longer scales with conversation length, five streaming tabs can be dragged simultaneously at 60fps, and idle CPU in background is near zero.

CLI and Toolchain

A few new commands stand out.

!command runs a shell command directly, at no model turn cost . like a quick terminal inside Hermes. /init scans a project and generates or updates your AGENTS.md. /diff shows staged, all, or session code changes. /context analyzes what is actually filling up your context window. /focus switches to a low-output view, with hidden content recoverable. hermes import-agent migrates Claude Code or Codex CLI configuration to Hermes in one step.

Mid-course redirects. If Hermes goes off track, you no longer need to /stop and re-explain everything. Just type a correction while it is working. The current turn redirects: work already done is preserved, the original prompt stays, and the agent adjusts course based on your new guidance. Paired with double-ESC discard and editor undo, correcting feels like editing, not restarting.

Tool self-healing. Truncated terminal output automatically overflows to a readable file and tells you the total size before truncation. The patch tool detects already-applied edits and returns success directly. write_file verifies actual disk content after writing. Zero-match searches auto-detect close alternatives. Default tool-call iteration cap raised from 90 to 500, so long autonomous tasks no longer hit an artificial ceiling.

Delegation and subagents. delegate_task gained structured timeout/stagnation metadata and real-time per-child state viewing via the /agents command. Subagents can now use execute_code. Subagent tool history is exposed in a sanitized form in subagent_stop, along with a public subagent lifecycle API for plugins.

Approvals and Security: Keeping Autonomy Controllable

The more capable an agent, the more control matters.

Smart approvals. hermes approvals suggest mines your approval history and auto-generates allowlist proposals: what you approved once becomes auto-approved next time. Operators can customize smart-approval policies. Consecutive rejections trigger a circuit breaker that freezes a misbehaving loop. Docker/podman daemon-redirect commands now require approval. Session-level runaway loops got caps on web_search and delegate_task, inspired by Claude Code.

Outbound security. The HMAC-signed outbound webhooks we covered earlier. DNS-pinned SSRF-safe fetch and Slack CDN allowlist prevent server-side request forgery. The credential pool, when hitting rate limits, stays on the fallback until the rate limit window resets . no immediate fail-back-and-fail-again loops. `${env:VAR}` SecretRef alignment between config.yaml and MCP config landed.

Windows hardening. An entire class of text-mode subprocess decoding bugs got closed across the entire repo. These bugs would never trigger on Linux . they only showed up in Windows production. Console flicker got hidden across all daemons, environment probes, LSP, and installer paths. Remaining encoding gaps (MCP stdio, gateway update I/O, STT/TTS, desktop spawn) got patched.

State and database integrity. Four session-state fixes landed (safe shutdown tracking, flush-cursor fixes, line retries, usage-PK healing). FTS layout upgraded to v23 with CJK bigram tokenization . Chinese search finally will not skip matches. Read paths got per-thread read-only connection isolation.

Compression and Performance: Long Conversations Stay Smooth, Cold Starts Are Fast

Context compression got a deep overhaul. Large-context models now proactively prune tool results instead of waiting until they fill up. Per-turn micro-compression spreads the cost across turns instead of one big pause. The last N user messages are guaranteed to survive (configurable via compression.min_tail_user_messages), so you will not see your most recent message get compressed away. Ghost-skill defense ensures pruned skills leave no phantom markers in the session. Thresholds can be configured per model and absolute token count.

Cold start and hot paths. hermes -w cold start dropped from about 14 seconds to about 1.8 seconds. hermes update no-op got 2 to 6 seconds faster. Heavy SDKs switched to lazy loading so they do not block the startup path. Config reading stopped doing deep copies . the telemetry gate got 54x faster and 29 call sites now use a read-only config loader, cutting read cost by 28x. Streaming hot loops dropped per-chunk repr() calls, making token stats roughly 3x cheaper. Turn flush batches into a single SQLite transaction. Anthropic native API prompt caching now covers tool schemas without losing history. DeepSeek got prompt caching added on the OpenCode gateway too.

MCP lazy start. Configured MCP servers no longer all load at session startup. A new fingerprint-keyed disk cache records each server is tool schema, and the corresponding server process only starts when needed. For people running many MCP servers, this is a real quality-of-life change: faster startup, less memory used.

FTS (full-text search) layout upgraded to v23 with CJK bigram tokenization. Chinese search finally will not miss matches.

Ecosystem Expansions

New platforms: Buzz (Nostr messenger), Vercel AI Gateway provider.

Platform upgrades: Slack native Block Kit clarify buttons, Discord auto-thread sessions keyed by prospective_thread_id, WhatsApp inbound read receipts now configurable, Photon native polls and rich links. Kanban wake now restores the creator is DM/thread session, with per-task model and thinking depth specifiable.

New models: Gemini 3.1 Pro and 3.6 Flash in the catalog (3.6-flash became the aux default), claude-opus-5 on OpenRouter and Nous Portal, deepseek-v4-flash-0731 available. Bedrock Converse API gained prompt caching.

Is It Worth Upgrading

Honestly, the codename Herald could not be more fitting.

v0.20.0 is the inflection point where Hermes goes from “typing assistant” to “talking, communicating, orchestratable agent platform.” Voice conversation is the most tangible change: Hermes can finally be interrupted, woken up, and talked to across the room . like talking to a person.

A2A, webhooks, and grounded citations together push Hermes into the infrastructure layer for multi-agent systems and event-driven workflows. The desktop crossed the platform threshold. The CLI picked up the power-user command suite people have been waiting for. Compression and cold starts stopped being things you notice and wait for.

3,650 commits in this window. The codename Herald is earned.