Source: Official GitHub release notes for tag v2026.8.3 (github.com/NousResearch/hermes-agent/releases/tag/v2026.8.3), fetched 2026-08-05, cross-checked against the announcement stub raw/x-account-nousresearch-2084325600643445095.md and two verbatim bookmark captures in raw/x-bookmarks-recent-digest-2026-08-05.md. Update via hermes update.
Provenance correction (2026-08-05). The staging stub for this release cautioned that “The Herald Release” looked like a model reconstruction, because the GitHub release title carries no codename. That caution was wrong — the release body opens with “The Herald Release.” The codename is first-party. The stub’s other unverified claims (voice wake-word, Buzz integration) also turned out correct, independently confirmed by verbatim bookmark captures.
v0.20.0 (released 2026-08-03) is the largest Hermes release the wiki has tracked, and it rolls up the v0.19.1 infrastructure patch tag (2026-07-30) entirely — so a vault sitting on v0.19.0 is two releases behind. The framing is literal: Hermes is the herald of the gods, and this release makes him one — he speaks (streaming conversational voice with barge-in and on-device wake words), carries word to other agents (A2A v1.0), announces events to your systems (signed outbound webhooks), and cites his sources (grounded research with verifiable citations).
Release by the numbers
Since v0.19.0: ~3,650 commits · ~1,400 merged PRs · ~5,200 files changed · ~559,000 insertions · ~405,000 deletions · ~1,200 issues closed · 650+ contributors.
Key Takeaways
- Voice became conversational, not transactional. Previously: speak, wait for the whole reply to generate, listen to one long audio file. Now Hermes synthesizes clause-by-clause as the response streams, you can interrupt mid-sentence just by talking (it stops, listens, and the model is told you cut in), and busy-aware silence detection stops it talking over you. Works in CLI voice mode, desktop, and gateway adapters.
- Wake words run on-device. Open-vocabulary wake phrase of your choosing; no audio leaves the machine while it waits. Multi-profile routing means different wake words reach different profiles, and saying “stop” ends voice chat on every surface. Off by default.
- Voice reaches the messaging platforms — voice notes on WhatsApp, Feishu, DingTalk, LINE, QQ, Photon, and Weixin are transcribed and answered, with platform-aware TTS delivery. STT is now fully configurable (own
hermes toolscategory, GUI toggles, unified language resolution that fixes the wrong-language-transcript class, OpenAIgpt-transcribesupport). grounded-citationsis the release’s most wiki-relevant skill. Every claim is backed by a verifiable source: quotes are matched against the actual page text rather than hallucinated, citations link to the exact evidence, and a fact-checking mode turns the same machinery on any document you hand it — reporting what checks out, what doesn’t, and what could not be verified. This is the verifier-first discipline shipped as a first-party feature.- Signed outbound webhooks end the polling requirement. Hermes now pushes lifecycle events (session activity, turn completions, tool events) to any HTTP endpoint you register, HMAC-signed so receivers can verify authenticity.
- **A2A v1.0 closes issue 514 — one of the oldest open feature requests in the repo. Hermes can now discover, talk to, and be driven by other Agent-to-Agent-compatible agents.
- The desktop app became a platform. Versioned artifacts with sandboxed live preview in a right-rail viewer, a real plugin SDK (Kanban is its founding plugin),
ctx.downloadfor handing users files, a global-hotkey quick-entry window, and multiple GUI windows. hermes import-agentmigrates a Claude Code or Codex CLI setup in one command — the most directly relevant line in the release for readers of Claude AI.- The tool-calling iteration limit jumped 90 → 500, removing an artificial ceiling on long autonomous runs.
The CLI power-user wave
| Command | What it does |
|---|---|
!command | Runs a shell command instantly without spending a model turn |
/init | Scans the project and generates or updates an AGENTS.md |
/diff | Staged / all / session changes from any surface |
/context | Breaks down exactly what is filling the context window |
/focus | Reduced-output view with hidden-line recovery |
Ctrl+S | Stashes a half-written prompt into a browsable panel |
hermes import-agent | Migrates a Claude Code or Codex CLI setup into Hermes |
Steering instead of restarting
Redirects let you correct the agent mid-turn: type a correction while it works and the active turn is redirected — work in flight is preserved, the original prompt is kept, and the agent course-corrects. Paired with double-ESC draft discard and a composer undo stack. The stated goal is that steering “feels like editing, not restarting.”
Tools that recover from their own failures
A sweep aimed at the turns agents waste on tool friction: truncated terminal output spills to a file the agent can read back; patch detects already-applied edits and diagnoses whitespace mismatches; write_file verifies content on disk; searches that match nothing probe for near-misses and recover; common failure classes return actionable hints.
Compression that respects the conversation
Proactive tool-result pruning for large-window models, per-turn micro-compaction that amortizes cost instead of one giant pause, a guaranteed N-user-message tail so recent conversation always survives (compression.min_tail_user_messages), progress-aware timeouts, and ghost-skill defense ([SKILL_PRUNED] markers) so a pruned skill cannot silently haunt a session. Thresholds are configurable per-model and in absolute tokens (compression.threshold_tokens).
Approvals and performance
hermes approvals suggestmines approval history into allowlist proposals; operators can customizeapprovals.smart_policy; a consecutive-denial circuit breaker stops a misbehaving loop cold; new approval gate for docker/podman daemon-redirect commands.hermes -wcold start dropped ~14s → ~1.8s; prompt caching now covers tool schemas on native Anthropic without history loss; config reads stopped deep-copying (54× faster on the telemetry gate); desktop shipped a second 60fps wave with streaming cost independent of transcript length.
New places to run and be reached
Buzz lands as a bundled gateway platform — Block’s Nostr-based messenger, with native WebSocket transport and NIP-42 auth. This is the packaged form of the integration announced separately on 2026-07-29, which described three ways to combine them: Buzz Desktop auto-discovering a local Hermes install, a relay bridge giving it a hosted identity in your channels, or connecting via the Hermes Gateway to use Buzz as a full external platform with channels, DMs, threads, reactions, and cron delivery. See Buzz. Also: the Vercel AI Gateway provider and Vercel Sandbox terminal backend return modernized, desktop gains an SSH remote-backend connection mode, and the Relay shipped four phases of parity (media, interactive prompts, thread lifecycle, typing indicators) plus HSP personal + org skill sync.
Community reception at ingest
- A community member built HermesCall, a proof-of-concept native iOS Swift app to “call” a Hermes agent, explicitly crediting “the changes in the latest release of Hermes (0.20.0)” — half-duplex, on-device PiperTTS, Tailscale IP as the phone number, not yet published to GitHub (
raw/reddit-1vg39kf.md). - The Kanban desktop plugin was demoed independently on X before the release notes landed, with a walkthrough running a real task through the Kanban process across multiple agent profiles.
Try It
hermes update— this is a two-release jump from v0.19.0.- Turn on a wake word and try hands-free CLI voice; detection is local and off by default.
- Run
/contexton a long session to see what is actually filling the window before reaching for compression settings. - If you maintain a Claude Code setup, try
hermes import-agentrather than porting configuration by hand. - Use the
grounded-citationsskill for any research task where “sounds right” is not good enough — then use its fact-checking mode on a document you already trust as a calibration test.
Open Questions
- Voice quality, latency, and wake-word false-positive rates are not quantified anywhere in the release notes — all descriptions are qualitative.
- The
grounded-citationsquote-matching mechanism is not specified. Exact string match, fuzzy match, or model-judged? This determines how much trust the “not hallucinated” claim earns. - v0.19.1 (
v2026.7.30) has no standalone article in this vault — it is described as an infrastructure patch tag rolled up here, but its contents were never separately reviewed. - Whether A2A v1.0 interoperates cleanly with Claude Code’s own multi-agent surfaces (subagents,
SendMessage) is untested here.