Emotion Chip

From its catalog listing
Your Hermes agent decides how to say it: the sighs, whispers and pauses it writes, voiced by ElevenLabs or Fish Audio. It writes its own delivery into the text it speaks, a small score in square brackets (a sigh, a whispered line, a pause, a warm and amused tone), and Emotion Chip renders exactly that for ElevenLabs v4 or Fish Audio (s2.1-pro, Drama 3), each in its own dialect, adding nothing the agent did not write. With voice replies on in a chat (/voice on), the chip tells the agent when Hermes will read its reply aloud, so the agent can say it itself, score included, instead of Hermes reading the plain text (not on Discord). Registers the emotion-chip speech provider, a delivery guide in the system prompt, the emotion-chip-voice skill, the /chip command (status says whether the chip can speak and what to fix; why, quiet, register) the hermes chip command, which shows offline what each engine receives for any score, and two agent tools: chip_status (read-only) and chip_preview (offline). CC0 emotion data and pre-registered evaluations ship in the repository.
Good to know before installing
with backend: eleven it sends your ElevenLabs API key, the rendered text, the model id, voice id, voice settings, an optional seed and, on v4 models, up to three earlier request ids from the same session (kept in memory for two hours) to api.elevenlabs.io over HTTPS; these are paid requests billed to your ElevenLabs account. The key is read at each call through the profile's secret scope and never stored. With the default backend: fish it sends nothing itself: it hands the rendered text to the Fish Audio plugin in the same process (which must be installed, enabled and signed in), and that plugin sends it to Fish Audio under its own disclosure and bills your Fish account; the chip reaches it through Hermes's speech-provider registry and imports its tag adapter. With guide_mode: generate (the default) it registers the emotion-chip-voice skill, and while it is the speech provider it adds a delivery guide of up to 4,000 characters to the system prompt and registers pre_gateway_dispatch (described next). With guide_mode: generate it also registers two agent tools. chip_status reads the plugin's settings, Hermes's tts configuration and this session's last authored register and reports them; it sends nothing and changes nothing. chip_preview renders a score offline and returns the text each engine would receive and the cues it would drop; no synthesis, no network, nothing written. While the chip is the speech provider, each successful text_to_speech result gains an emotion_chip entry of at most 600 characters (the dialect, the agent's cue words, drops and notes); it stays in the conversation and is sent nowhere. The pre_gateway_dispatch hook: for each incoming message it asks the gateway whether it will read the reply aloud and keeps that answer in memory under the chat and message id (at most 256, never used after 15 minutes), used once to add one line to that message's turn; it never skips or changes a message, and registering the hook routes webhook and relay deliveries through Hermes's hooked ingress path. When another speech provider is selected, its tool_request middleware rewrites text_to_speech text containing a bracket for that provider (the Fish dialect on Fish S2 and Drama 3 models, or the ElevenLabs dialect on a v3/v4 model; for any other provider or model, including Hermes's default Edge, Fish S1 and ElevenLabs on older models, standalone bracketed words are removed as cues, a bracketed Step 2 too, while the words inside spans, numeric references such as [1] and code such as an array index stay), leaves command providers alone and records nothing. With signals on (off by default) it adds a dated situation line of up to 200 characters to the turn and, when a tool call fails, to that call's result. It reads Hermes's tts configuration (read-only) to pick the dialect, model and voice. It writes under the profile's plugin-data/emotion-chip/ only the generated skill and, unless delivery_log is off, delivery.jsonl (time, session id, sender id, platform, model, dialect, rendering group, cue words, parameters, kinds of drops and character count, never the spoken text), moved to delivery.1.jsonl at 5 MB so at most about 10 MB is kept; Hermes keeps its per-session state (task counters and timestamps, the ids of the last 16 tool calls counted, session keys, the messaging sender id and platform, the last situation and register lines, the quiet switch; the 50 most recent sessions) in the plugin's own namespaced folder under plugin-data/. Audio goes to the path Hermes supplies. No background processes, no shell commands, no telemetry.
From its catalog listing.
How to install
- Install it. Hermes asks for any API keys it needs (this one needs
ELEVENLABS_API_KEY), then asks whether to turn it on.hermes plugins install emotion-chip - Said no? Turn it on later:
hermes plugins enable emotion-chip - Restart Hermes so it loads the plugin.
Some plugins also need a sign-in or settings: see Full setup & settings. Memory plugins ask whether to become your memory provider instead. Steps follow the official plugins guide (opens in a new tab).
- Category
- Chat apps & voice
- Maintained by
- 100yenadmin (Community)
- Needs
ELEVENLABS_API_KEYAPI key- Adds tools
chip_statuschip_preview- Adds hooks
pre_llm_callpost_tool_callapi_request_erroron_session_starton_session_endtransform_tool_resulton_session_resetpre_gateway_dispatch- Works on
- linux, macos
- Needs Hermes
- 0.22.2 or newer
- Version
- 0.3.3
- In the catalog
- 2026-11-07, updated 2026-11-11
Known issues (from its catalog listing)
- Full cues reach the voice on the text_to_speech tool path, and in chats with voice replies on, where the agent says its reply itself. Not on Discord, where Hermes plays its own reply in voice channels. Read-aloud and voice mode speak the reply text without cues, because Hermes has no seam yet that keeps cues for speech and strips them from the chat.
- The voice-replies line needs the plugin to run inside the gateway process (Hermes's default); an isolated plugin host gets no line, which is Hermes's behaviour without the chip.
- The Fish backend needs the fish-audio plugin installed, enabled and signed in.
- ELEVENLABS_API_KEY is needed only with backend: eleven; the default Fish backend uses the fish-audio plugin's key.
- GPT-Live voice chat speaks through OpenAI's live voice model, not a speech provider, so the chip does not reach it.
Using the Hermes desktop app? See where plugins live

Full setup & settings ↗Source code ↗
More in chat apps & voice
New Hermes Agent updates every day on X. Follow @HermesWatcher