hermes-agent: deprecate

[BrewTestBot]

Created by brew bump

---

Created with brew bump-formula-pr.

• resource blocks have been checked for updates.

release notes

# Hermes Agent v0.20.0 (v2026.8.3)
Release Date: August 3, 2026

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

The Herald Release. Hermes is the herald of the gods, and this release makes him one in earnest: he speaks (real-time conversational voice with streaming TTS, barge-in, on-device wake words, and hands-free control across the CLI, desktop, and every audio-capable gateway platform), he carries word to other agents (A2A v1.0), he announces events to your systems (signed outbound webhooks), and he cites his sources (grounded research with verifiable citations and fact-checking). Around that spine: the desktop app became a platform (artifacts with live preview, a plugin SDK, quick-entry from anywhere, multiple windows), the CLI got a wave of power commands (! shell mode, /init, /diff, /context, /focus), compression got smarter and gentler, and the tools themselves now recover from their own failures instead of making the model guess. This release rolls up everything from the v0.19.1 infrastructure patch tag — that window is fully documented here.


✨ Highlights


Talk to Hermes — streaming, conversational voice with barge-in — Voice mode used to mean: speak, wait for the whole reply to generate, then listen to one long audio file. Now Hermes speaks clause-by-clause as the response streams, you can interrupt it mid-sentence by just talking (it stops, listens, and the model is told you cut in), and busy-aware silence detection means it doesn't talk over you. This works in CLI voice mode, on the desktop, and through gateway adapters. Talking to Hermes finally feels like a conversation, not a voicemail exchange. (#69511, #73862, #74223, #74000, #69602 — @teknium1, @OutThisLife)


Wake words and hands-free control — Say your own open-vocabulary wake phrase ("hey Hermes", or anything you pick) and Hermes starts listening — detection runs on-device, so no audio leaves your machine while it waits. Multi-profile voice routing means different wake words can reach different profiles, and saying "stop" ends the voice chat on every surface without touching the keyboard. Your terminal is now something you can talk to from across the room. (#70509, #73106, #73933 — @teknium1)


Voice on every platform — Send a voice note to Hermes on WhatsApp, Feishu, DingTalk, LINE, QQ, Photon, or Weixin and it's transcribed and answered; auto-TTS replies are delivered platform-aware (opus where platforms want opus, captions attached correctly). STT is now fully configurable — its own hermes tools category, GUI toggles, dashboard dropdowns, unified language resolution so transcripts stop coming back in the wrong language, and OpenAI's gpt-transcribe support. One unified spoken-text preprocessor cleans markdown, code, and URLs out of speech across all TTS providers. (#73515, #73508, #73910, #73513, #73067 — @teknium1)


Research you can trust — grounded citations with fact-checking — The new grounded-citations skill makes Hermes produce research where every claim is backed by a verifiable source: quotes are matched against the actual page text (not hallucinated), citations link to the exact evidence, and a fact-checking mode turns the same machinery on any document or claim you hand it — it tells you what checks out, what doesn't, and what couldn't be verified. If you use Hermes for research, this is the difference between "sounds right" and "provably sourced." (#71698, #77104 — @teknium1)


Outbound webhooks — Hermes pushes events to your systems — Until now, integrating with Hermes meant polling or listening on a platform. Now Hermes pushes signed lifecycle events (session activity, turn completions, tool events) to any HTTP endpoint you register — with HMAC signatures so your receiver can verify authenticity. Wire Hermes into your CI, your home automation, your dashboards, or any service that speaks HTTP, with no polling loop. (#69406 — @teknium1)


The desktop app becomes a platform — artifacts, plugin SDK, quick entry — Hermes desktop now renders artifacts: versioned cards with sandboxed live preview in a right-rail viewer, so generated HTML/apps run safely next to the chat. A real plugin SDK landed with Kanban as its founding plugin, ctx.download for handing users files, floating pane placement, and multiple GUI windows. A global-hotkey quick-entry window captures a thought into any session from anywhere in your OS. The desktop stopped being a chat client and started being a workbench. (#72345, #61173, #74413, #72315, #68259, #73143 — @OutThisLife, @teknium1)


Hermes speaks Agent-to-Agent — A2A v1.0 — A new bundled plugin implements the Agent-to-Agent protocol, so Hermes can discover, talk to, and be driven by other A2A-compatible agents. This closes issue #514 — one of the oldest open feature requests in the repo. If you're building multi-agent systems with heterogeneous stacks, Hermes now has a standard wire protocol for joining them. (#77109 — @teknium1)


CLI power-user wave — !command runs a shell command instantly without spending a model turn. /init scans your project and generates (or updates) an AGENTS.md. /diff shows staged/all/session changes from any surface, /context breaks down exactly what's filling your context window, /focus gives you a reduced-output view with hidden-line recovery, and Ctrl+S stashes a half-written prompt into a browsable panel. Plus hermes import-agent migrates your Claude Code or Codex CLI setup into Hermes in one command. (#72257, #72178, #72240, #72242, #72302, #72262, #72190 — @teknium1, several salvaging long-standing community PRs)


Correct the agent mid-turn — redirects — If Hermes is heading the wrong way, you no longer have to /stop and re-explain. 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 with your new guidance. Paired with double-ESC draft discard and a composer undo stack, steering feels like editing, not restarting. (#63104, #72339, #74736 — @OutThisLife)


Tools that fix themselves — A sweep of self-recovery upgrades means the agent wastes far fewer turns 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, and common failure classes come back with actionable hints. The default tool-calling iteration limit also jumped 90 → 500 — long autonomous runs stopped hitting an artificial wall. (#77041, #76998, #77024, #77055, #77011, #76992, #72176 — @teknium1)


Compression that respects your conversation — Context compression got a deep overhaul: proactive tool-result pruning for large-window models, per-turn micro-compaction that amortizes the cost instead of one giant pause, a guaranteed N-user-message tail so recent conversation always survives, progress-aware timeouts that stop punishing slow summary models, and ghost-skill defense so a pruned skill can never silently haunt a session. Thresholds are now configurable per-model and in absolute tokens. Long sessions stay coherent and stop stalling. (#70254, #75345, #70250, #71508, #70275 — @teknium1, @kshitijk4poor, salvaging multiple community PRs)


Smart approvals grow up — hermes approvals suggest mines your approval history into allowlist proposals, operators can customize the smart-approval policy, a consecutive-denial circuit breaker stops a misbehaving loop cold, and desktop pairing approvals are profile-correct with a proper surface to answer them from. Plus a new approval gate for docker/podman daemon-redirect commands. Less clicking "approve", without giving an inch of control. (#72259, #72186, #72203, #74446, #71092 — @teknium1, @OutThisLife)


Faster everywhere, again — Prompt caching now covers tool schemas on native Anthropic without history loss. hermes -w cold start dropped ~14s → ~1.8s, hermes update no-ops got 2–6s faster, heavy SDKs lazy-load off the import path, config reads stopped deep-copying (54× faster on the telemetry gate), and the desktop shipped a second 60fps wave — streaming cost independent of transcript length, drag at 60fps with five streaming tabs, idle CPU near zero in the background. (#76032, #71637, #74218, #74204, #71835, #72346, #75218 — @kshitijk4poor, @teknium1, @OutThisLife)


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), 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. (#73610, #73761, #74518, #68130, #71300–#71624, #66730 — @teknium1, @yoniebans, @benbarclay)



🎙️ Voice & Speech
Conversational voice

Streaming, conversational TTS with barge-in across all surfaces; clause-by-clause synthesis for CLI voice mode + gateway adapters (#69511, #73862 — @OutThisLife, @teknium1)
Voice chat UX polish — busy-aware silence, stop hint, thinking sounds, barge-in fix; full-duplex turn listener (interrupt by voice during generation AND playback) (#74000, #74223 — @teknium1)
On-device wake words with open-vocabulary phrases + multi-profile voice routing; say "stop" to end voice chat hands-free on every surface (#70509, #73106, #73933 — @teknium1)
The model is told when the user interrupts its spoken reply; desktop speaks the whole turn and idle-flushes held narration (#69602, #69936 — @OutThisLife, @teknium1)
15-item CLI/TUI voice-mode UX and environment fix wave (#73520 — @teknium1)

TTS / STT infrastructure

Unified spoken-text preprocessing + speed/instructions/provider tool params; unified STT language resolution (fixes the wrong-language transcription class); global stt.language defaults to en (#73513, #73067, #73100 — @teknium1)
Fully configurable STT — hermes tools category, GUI toggle/matrix, dashboard dropdowns, setup status; OpenAI gpt-transcribe support (#73910, #73853 — @teknium1)
Platform-aware auto-TTS voice delivery (opus platforms, streamed/global gap, captions); inbound voice classification/routing for Feishu, DingTalk, LINE, QQ, Photon, WhatsApp, Weixin (#73508, #73515 — @teknium1)
Command TTS/STT provider hardening — idle timeouts, env scrubbing, no-shell, path guards (#73514 — @teknium1)
Sync per-sentence TTS synthesis pipelined with playback — the next sentence renders while the current one speaks (#77355 — @kshitijk4poor)
Discord voice PCM streams to ffmpeg stdin instead of a temp file (#76970 — @kshitijk4poor)

🏗️ 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 input with head+tail retention (#70254, #75345, #70250, #70249 — @teknium1, @kshitijk4poor)
Ghost-skill defense — [SKILL_PRUNED] markers, protected prune, deterministic survival; progress-aware timeouts; lock-contended compression soft-defers instead of exhausting (#70275, #71508, #70285 — @teknium1)
Per-model threshold overrides; absolute token threshold (compression.threshold_tokens); opt-in idle-triggered compaction; opt-in progress notices; structured local logging for compression attempts (#69339, #69335, #69360, #70457, #69338 — @teknium1)
Context-engine ABC grows select_context() + on_turn_complete() verbs (salvage of @chaos-xxl's RFC work); engines can suppress or customize compaction status (#70458, #69859 — @teknium1)
Strict redaction applied at every compaction text boundary (#69294 — @teknium1)

Prompt caching & hot-path performance

Tool schemas cached on native Anthropic without history loss + consolidated cache-plan internals (#76032, #76067 — @kshitijk4poor)
DeepSeek prompt caching on OpenCode gateways; per-API-call token accounting off the turn thread; OpenAI wire client reused across sequential LLM calls; send-path tool-call canonicalization memoized (#75886, #73359, #73375, #76880 — @teknium1, @kshitijk4poor)
Readonly config loader at 29 call sites (28× cheaper reads); per-turn config deepcopies killed (telemetry gate 54×); one raw config.yaml parse per process; inter-tool delay removed (#74322, #74211, #74228, #64172 — @teknium1, @Soju06)
Lazy heavy-SDK imports (−8-10% import cost on top of the mcp/tool-discovery diet); streaming hot loop drops per-chunk repr() (~3× cheaper accounting); cursor/memo optimizations for per-iteration history walks (#74204, #74194, #74221, #74231 — @teknium1)
Cold-start ~14s GIL stall during backend init mitigated; turn flush batched into one SQLite transaction; provider-capability-gated prompt cache keys (implied for api.openai.com) (#77814, #77619, #77609 — @kshitijk4poor)
AIAgent hot-path salvage — prompt-cache copy, reasoning-timeout precompute, lazy compressor init (#57229 — @kshitijk4poor)

Approvals & the agent loop

hermes approvals suggest mines approval history into allowlist proposals; operator-customizable approvals.smart_policy; consecutive-denial circuit breaker; cross-surface approvals mode command (#72259, #72186, #72203, #63517 — @teknium1)
Docker/podman daemon-redirect commands require approval; session-wide runaway-loop caps for web_search + delegate_task (Claude Code-inspired) (#71092, #66600 — @teknium1)
Mid-turn redirects — user corrections steer the active turn, preserving in-flight work and the original prompt (#63104, #72339 — @OutThisLife)
Delegation: structured timeout/stall metadata + live per-child status in /agents; subagents can use execute_code; redacted child tool history exposed in subagent_stop; public subagent lifecycle API for plugins (#72300, #69325, #72403, #72501 — @teknium1)
Single-owner refactors for backend identity + failure-scoped skips, empty-content wire repair, call_id/reasoning sanitization, model-switch parsing (#72505, #73071, #74319, #74229 — @teknium1)
Labeled reasoning excerpt surfaced at the empty-response terminal; tool_search probe-validates blind tool_call args (#65144, #59267 — @teknium1)

Tool self-recovery wave

Terminal: recoverable truncation (full output spilled + pre-truncation size), cwd echoed when a command changes directory, output-pattern failure hints (#77041, #77004, #76992 — @teknium1)
Patch: already-applied edits return success no-op, whitespace-visualized no-match diagnosis, ambiguous-match locations listed (#76998, #77024, #77001 — @teknium1)
Search: zero-match probes + multi-path recovery, auto-multiline for newline patterns; read_file default limit 500 → 2000 lines; negative-result cache for read/search misses; write_file verifies on-disk content (#77011, #77102, #76996, #76945, #77055 — @teknium1, @kshitijk4poor)
execute_code recovery hints; skill_view dedup stub for unchanged re-reads; terminal/execute_code schema prose trimmed ~40%; tiered tool disclosure scales with catalog size; default iteration limit 90 → 500 (#77106, #77095, #77023, #67034, #72176 — @teknium1)

Providers & models

Vercel AI Gateway provider + Vercel Sandbox terminal backend return, modernized (SDK 0.7.2, telemetry off) (#74518 — @teknium1)
Gemini 3.1 Pro + 3.6 Flash in catalogs; Gemini salvage cluster (3.6-flash aux default, Vertex catalog, direct cost tracking); claude-opus-5 in OpenRouter + Nous Portal; deepseek-v4-flash-0731 (#73479, #73516, #70946, #75501 — @teknium1)
Bedrock Converse API prompt caching (cachePoint) (#70231 — @JoaoMarcos44)
OpenAI data-residency endpoints get declared transport + correct catalog; provider-aware API-server request routing; backend-acknowledged session model lock; Nous sticky routing via top-level session_id (#74958, #70853, #70950, #69253 — @victor-kyriazakos, @teknium1)
Model picker: curated defaults + collapsible providers + select-all; stale caches served instantly with background refresh; custom-endpoint probe capped at 1.5s; honcho OAuth device-code login (#73172, #76430, #76922, #61608 — @OutThisLife, @teknium1, @kshitijk4poor, @akattelu)
ACP: named custom providers in the model selector; authenticated cross-provider model choices; non-blocking startup via background MCP discovery (#70082, #70404, #75985 — @israellot, @amanning3390, @kshitijk4poor)

Secrets & config

Command-helper secret source (composes with all vaults); one-command token rotation + actionable startup errors; opt-in encrypted break-glass cache for Bitwarden; vault-injected keys scoped per profile home; orchestrator preserve_existing + profile aliasing (#69266, #68605, #69251, #69250, #69058 — @teknium1)
${env:VAR} SecretRef parity between config.yaml and MCP config; secret-source env vars reach stdio MCP servers (#69267, #69053 — @teknium1)
Canonical config loaders for behavioral reads; table-driven config migration registry; DEFAULT_CONFIG extracted to config_defaults.py; auto-migration support floor at v12 (#74237, #74200, #74182, #74433 — @teknium1)

🌐 Gateway, Relay & Fleet

Session activity heartbeats, stall watchdog, and bounded compression waits — re-landed hardened after an in-window revert cycle (originally #72424 by @fangliquanflq) (#76354 — @teknium1)
SessionState consolidation (19 session-keyed dicts → one turn/conversation/persistent-scoped object); TurnContext/TurnRunner seam extraction; declarative busy_policy on CommandDef (#74289, #74353, #74197 — @teknium1)
Relay parity waves: Phase 1 (supported_ops discovery, identity fields, /handoff aliasing), Phase 2 media, Phase 3 interactive prompts, Phase 4 thread lifecycle; egress typing indicators (#71300, #71363, #71404, #71624, #69721 — @benbarclay)
HSP skill sync: personal client (M1) + org-skills client (M2) + org-skill namespace with token-gated discovery (#66730, #70024, #70459 — @benbarclay)
Buzz (Block/Nostr) platform adapter with native WebSocket inbound transport + NIP-42 auth (#73610, #73761 — @teknium1)
Photon: native polls, effects, clarify-as-poll, rich links (4-PR salvage) (#73614 — @teknium1)
Slack: native Block Kit clarify buttons; opt-in reaction triggers; outbound payload sanitization; thread-context lifecycle fixes (#69318, #70195, #69317, #69320 — @teknium1)
Discord auto-thread sessions keyed on prospective_thread_id; reply references built from ids (no fetch_message); WhatsApp configurable inbound read receipts (#76513, #76875, #73322 — @benbarclay, @kshitijk4poor)
Kanban wakes resume the creator's DM/thread session; kanban/delegate wake-ups reach api_server sessions; per-task model + thinking-depth from the board (#72191, #70171, #69876, #76417 — @teknium1, @OutThisLife)
Relay: Discord tool-progress routed into the auto-thread instead of the parent channel (#77830 — @benbarclay)
Outbound webhooks — push signed lifecycle events to external endpoints; simplex channel enumeration in hermes send --list (#69406, #77110 — @teknium1)

🖥️ Hermes Desktop App
The platform wave

Artifacts — versioned cards, sandboxed live preview, right-rail viewer (#72345 — @teknium1)
Plugin SDK — Kanban as the founding desktop plugin; ctx.download hands the user a file; widget-app SDK (apps as state+reducer+render) with three reference apps; widget-grid layout engine + background-aware theme engine (#61173, #74413, #68306, #20379 — @OutThisLife)
Quick-entry window (global hotkey → any session); multiple GUI windows; floating pane placement; pane toggles anywhere + hidden header; ⌘O open-folder-as-project (#72315, #68259, #73143, #75848, [#746...

View the full release notes at https://github.com/NousResearch/hermes-agent/releases/tag/v2026.8.3.

---

[iMichka]

To be able to build with Python 3.14, this line was added:

# Allow to build with Python 3.14

inreplace "pyproject.toml", "requires-python = ">=3.11,<3.14"", "requires-python = ">=3.11,<3.15""

When our CI runs, "brew update-python-resources hermes-agent" fails with:

ERROR: Package 'hermes-agent' requires a different Python: 3.14.6 not in '<3.14,>=3.11'

Because the inreplace is not executed to update dependencies.

I have edited the formula locally to make it use Python 3.13, ran the update manually and committed that change.
It's not perfect, but it should work.

Note sure if we should work on a better workaround ...

[botantony]

NousResearch/hermes-agent#68217

[iMichka]

Deprecated then. If upstream ripped out support, and we are stuck on an old Python version, we will need to move on.

Deprecation phase is 6 months, as per policy https://docs.brew.sh/Deprecating-Disabling-and-Removing this formula has more than 2000 downloads in the last 30 days, so we need time to give people some notice.

[daeho-ro]

• hermes-agent 2026.7.30 #296243

[iMichka]

Undeprecated in #297327 as a solution was found to still build it.