Skip to content

Hermes Agent v0.20.0 (v2026.8.3)

Release Date: August 3, 2026 Since v0.19.1: ~945 commits · ~300 files changed · 82 contributors

The Herald Release. Voice becomes conversational, Hermes talks to other agents, research gets citations you can verify, the desktop becomes a platform, and the CLI gets a power-user upgrade.

✨ Highlights

  • Streaming Conversational Voice with Barge-In — Clause-by-clause TTS synthesis, interrupt by voice mid-sentence, busy-aware silence detection. Works across CLI voice mode, desktop, and gateway adapters. Talking to Hermes finally feels like a conversation.

  • Wake Words & Hands-Free Control — Open-vocabulary wake phrases ("hey Hermes" or anything you pick), on-device detection (no audio leaves your machine), multi-profile voice routing. Say "stop" to end voice chat hands-free on every surface.

  • Voice on Every Platform — Voice notes on WhatsApp, Feishu, DingTalk, LINE, QQ, Photon, Weixin transcribed and answered with platform-aware auto-TTS replies. Fully configurable STT with its own hermes tools category, GUI toggles, and OpenAI gpt-transcribe support.

  • Grounded Citations & Fact-Checking — The new grounded-citations skill matches quotes against actual page text (not hallucinated), links citations to exact evidence, and a fact-checking mode verifies any document or claim. Research goes from "sounds right" to "provably sourced."

  • Outbound Webhooks — Hermes pushes signed lifecycle events (session activity, turn completions, tool events) to any HTTP endpoint with HMAC signatures. Wire Hermes into CI, home automation, dashboards, or any HTTP-speaking service — no polling loop.

  • Desktop Becomes a Platform — Artifacts (sandboxed live preview in right-rail viewer), plugin SDK (Kanban as founding plugin, ctx.download, floating panes, multiple GUI windows), global-hotkey quick-entry window. The desktop is now a workbench, not just a chat client.

  • Agent-to-Agent Protocol (A2A v1.0) — Bundled plugin implementing the A2A standard. Hermes can discover, talk to, and be driven by other A2A-compatible agents. Closes issue #514 — one of the oldest feature requests.

  • CLI Power-User Wave — !command runs shell instantly without a model turn. /init generates AGENTS.md. /diff shows staged/all/session changes. /context breaks down context window usage. /focus enables reduced-output view. Ctrl+S stashes prompts. hermes import-agent migrates from Claude Code/Codex CLI in one command.

  • Mid-Turn Redirects — Correct the agent while it works — the active turn is redirected with your new guidance, preserving in-flight work and the original prompt. Double-ESC discards drafts, composer undo stack built in.

  • Tools That Fix Themselves — Truncated terminal output spills to a readable file. patch detects already-applied edits and diagnoses whitespace mismatches. Searches probe near-misses and recover. write_file verifies on-disk content. Table stakes for long autonomous runs. Default tool iteration limit raised from 90 to 500.

  • Smarter, Gentler Compression — Proactive tool-result pruning, per-turn micro-compaction, guaranteed N-user-message tail survival, progress-aware timeouts, ghost-skill defense. Long sessions stay coherent and stop stalling.

  • Smart Approvals Grow Up — hermes approvals suggest mines history into allowlist proposals. Operator-customizable smart-approval policy. Consecutive-denial circuit breaker stops misbehaving loops cold. Desktop pairing approvals are profile-correct.

  • Faster Everywhere, Again — Prompt caching covers tool schemas on native Anthropic. hermes -w cold start: ~14s → ~1.8s. Config reads 54× faster. Desktop 60fps wave 2: streaming cost independent of transcript length, idle CPU near zero.

🎙️ Voice & Speech

  • Streaming conversational TTS with barge-in across all surfaces; clause-by-clause synthesis
  • Voice chat UX: busy-aware silence, stop hint, thinking sounds, full-duplex turn listener
  • On-device wake words with open-vocabulary phrases + multi-profile voice routing
  • Model is told when the user interrupts its spoken reply
  • Unified spoken-text preprocessing + unified STT language resolution (fixes wrong-language transcription)
  • Fully configurable STT: hermes tools category, GUI toggle, dashboard dropdowns, OpenAI gpt-transcribe
  • Platform-aware auto-TTS voice delivery (opus platforms, streamed gaps, captions)
  • Inbound voice classification/routing for Feishu, DingTalk, LINE, QQ, Photon, WhatsApp, Weixin
  • Sync per-sentence TTS pipelined with playback; Discord voice PCM streams to ffmpeg stdin

🏗️ Core Agent & Architecture

Compression & Context

  • Proactive tool-result pruning for large-window models; per-turn micro-compaction
  • N-user tail guarantee (compression.min_tail_user_messages); bounded summarizer with head+tail retention
  • Ghost-skill defense; progress-aware timeouts; lock-contended compression soft-defers
  • Per-model threshold overrides; absolute token threshold; opt-in idle-triggered compaction
  • Strict redaction at every compaction text boundary

Prompt Caching & Performance

  • Tool schemas cached on native Anthropic without history loss
  • DeepSeek prompt caching on OpenCode gateways; per-API-call token accounting
  • Readonly config loader at 29 call sites (28× cheaper reads); per-turn deepcopies killed (telemetry gate 54×)
  • Lazy heavy-SDK imports (−8-10% import cost); streaming hot loop ~3× cheaper accounting
  • Cold-start ~14s GIL stall mitigated; turn flush batched into one SQLite transaction

Approvals & Agent Loop

  • hermes approvals suggest mines approval history into allowlist proposals
  • Consecutive-denial circuit breaker; docker/podman daemon-redirect approval gate
  • Mid-turn redirects — user corrections steer the active turn
  • Delegation: structured timeout/stall metadata, live per-child /agents status, subagent execute_code access
  • Default tool iteration limit 90 → 500; tool_search validates blind tool_call args

Tool Self-Recovery Wave

  • Terminal: recoverable truncation (full output spilled to file), cwd echoed on directory change, failure hints
  • Patch: already-applied edits return success no-op, whitespace-visualized diagnosis, ambiguous-match locations
  • Search: zero-match probes + multi-path recovery, auto-multiline for newline patterns
  • read_file default limit 500 → 2000; negative-result cache for misses; write_file verifies on-disk content
  • execute_code recovery hints; skill_view dedup stub; tiered tool disclosure for large catalogs

Providers & Models

  • Vercel AI Gateway provider + Vercel Sandbox terminal backend return (SDK 0.7.2)
  • Gemini 3.1 Pro + 3.6 Flash in catalogs; claude-opus-5 in OpenRouter + Nous Portal; deepseek-v4-flash-0731
  • Bedrock Converse API prompt caching (cachePoint)
  • OpenAI data-residency endpoints get declared transport + correct catalog
  • MCP lazy server startup from fingerprint-keyed on-disk tool-schema cache

Secrets & Config

  • Command-helper secret source (composes with all vaults); one-command token rotation
  • ${env:VAR} SecretRef parity between config.yaml and MCP config
  • Canonical config loaders + table-driven migration registry; auto-migration support floor at v12

🌐 Gateway, Relay & Fleet

  • Session activity heartbeats, stall watchdog, bounded compression waits
  • SessionState consolidation (19 session-keyed dicts → one scoped object)
  • Relay parity waves: media, interactive prompts, thread lifecycle; egress typing indicators
  • HSP skill sync: personal + org-skills client with token-gated discovery
  • Buzz (Block/Nostr) platform adapter with native WebSocket transport + NIP-42 auth
  • Photon: native polls, effects, clarify-as-poll, rich links
  • Slack: native Block Kit clarify buttons; opt-in reaction triggers
  • Discord auto-thread sessions; WhatsApp configurable inbound read receipts
  • Outbound webhooks — push signed lifecycle events to external endpoints

🖥️ Desktop App — The Platform Wave

Artifacts, Plugins & Quick Entry

  • Artifacts — versioned cards, sandboxed live preview, right-rail viewer
  • Plugin SDK — Kanban founding plugin, ctx.download, widget-app SDK, widget-grid layout engine
  • Quick-entry window (global hotkey → any session); multiple GUI windows; floating pane placement
  • SSH remote-backend connection mode; event-driven live sync replaces always-on polls
  • Agent can drive the shell (preview pane + pane focus) AND inspect the desktop app it's developing

Composer & UX

  • Attach files/folders/links via picker; composer chips for @paths and pasted links
  • Composer undo stack; double-ESC discards draft; double-Enter sends; type-to-focus
  • 2-keypress model switching (⌘⇧M); YOLO in ⌘K with live toggle; keyboard-first pickers
  • Grouped, live-ticking tool-activity line; @session links resolve to clickable titles
  • iMessage-style emoji reactions (opt-in, two-way); double-click to heart
  • Sidebar date dividers + pinned section; credit-usage toasts; Cron Blueprints + Webhooks pages

Desktop Performance (60fps Wave 2)

  • Streaming cost independent of transcript length; 60fps on real sessions
  • Drag at 60fps with five streaming tabs; multitab streaming fast
  • Idle CPU near zero in background; sidebar/overlay render churn killed
  • ⌘K opens instantly; renderer cold start keeps shiki/mermaid off the boot path
  • Playwright E2E suite with visual regression diffs

🖥️ CLI, TUI & Dashboard

  • ! shell mode; /init AGENTS.md generation; /diff (staged/all/session); /context breakdown; /focus reduced-output
  • Ctrl+S prompt stash; persistent /goal indicator; multi-select clarify across CLI/gateway/TUI
  • hermes import-agent — one-command migration from Claude Code/Codex CLI
  • Per-turn summary line + live token flow in spinner; cross-surface theme SDK
  • hermes -w startup ~14s → ~1.8s; banner update-check 6× faster
  • Dashboard lazy-loads routes + GROUP BY session stats; session filtering tabs
  • Arabic (ar) locale with RTL across desktop/dashboard/agent

🧩 Skills, Plugins & MCP

  • A2A v1.0 agent protocol plugin (closes #514)
  • Curator: surface unmanaged skills + curator adopt; grounded-citations skill + fact-checking mode
  • Office skills bundled: docx, xlsx, pdf + refreshed powerpoint
  • Skills-tree debloat (yuanbao, segment-anything, jupyter → optional-skills)
  • MCP: Comfy Cloud catalog entry with curated 20-tool default; lazy server startup from tool-schema cache
  • NeMo Relay observability integration

🔒 Security & Reliability

  • Iron-proxy credential-injection egress firewall
  • DNS-pinned SSRF-safe fetches + Slack CDN allowlist; ReDoS eliminated in config-key redaction
  • Tier-3 credential reads scoped; CVE dependency pins refreshed
  • Windows hardening: text-mode subprocess decode bug class closed, console flashes hidden
  • Four session-state fixes; compact v23 FTS layout + hermes sessions optimize
  • OpenViking memory-provider hardening — fail closed on blocked endpoints

👥 Contributors

647 contributors shipped this release.

Core: @teknium1, @OutThisLife (desktop, voice, perf), @kshitijk4poor (perf, caching, salvage), @ethernet8023 (runtime, E2E, desktop), @benbarclay (relay, HSP, auth)

Thank you to all 647 contributors who contributed code, co-authored fixes, filed the ~1,200 issues this release closes, and had their PRs salvaged into main.

Updating

hermes update
# or fresh install:
curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash

Full Changelog: v2026.7.20...v2026.8.3

← v0.19.1 — Patch Release | Changelog Home →

↑ Changelog Home

This Hermes repo is one of the largest structured collections of public AI, automation, business, and technology documentation. Content remains attributed to original authors and repositories. Indexed and organized by www.CorpusIQ.io.