Myna settings

Role

Default role (also in the panel) Track topics and build the profile and conversation map Show all roles (business development, partnerships, support, interviews and more), not only Sales, Investor and Listen Lab mode: entertainment roles (Stand-up, Stand-up roast, Flirty chat, Active listening) and partner-change detection for chat roulettes

Edit role instructions

Role instructions: every role can be edited

Applies to the next card. The header (what the role shows, bridges) stays as shipped; the text is what the model reads. USER is you, OTHER is the person on the call.

Custom role instructions Typed decisions (TypeSafe Jev, experimental)

Knowledge

Where the facts on cards come from. Pick any combination.

Knowledge base: your documents and pasted notes. Marked with a filled square.

Relevance threshold (lower: more facts, more noise) Facts per card

Add material by pasting Knowledge set

The model splits the text into facts and stories. They arrive as drafts: check them and press Verify.

Raw notes without slicing (sent as is, only while the knowledge base is empty)

Import a product page

Reads the page and linked pricing and feature pages, adds facts and a sales script to the knowledge base as drafts.

Model knowledge: what the model already knows: definitions, public companies, history. Marked with a hollow circle. Can be outdated or wrong; never used for your own numbers. Web search: current facts with a link to the source. Marked with a dashed square and the site name.

When to search Search engine

Once per topic: cards keep their speed, web facts appear a few seconds after the topic changes. On every question: the freshest facts, but each card waits for the search.

Results per search

The Investor pitch role always uses the knowledge base only, whatever is selected here.

Cards

Show cards automatically when the other person speaks Follow the topic: a new card when the topic changes, refreshed during long stretches without questions (needs topic tracking in Role)

What a card shows (same as the three buttons in the panel)

Facts: short, checkable items: numbers, names, dates, conditions. You glance at them and say it your way. Advice: one line on what the question is really about and where to take the conversation. Not meant to be read out. Answer: a finished reply in the first person, up to 40 words, built from the facts. Takes a little longer.

“Is there a discount if we pay for the year?”

Facts25% off with yearly billing: $6 instead of $8 · not combined with the startup program

AdvicePrice-sensitive. Give the yearly figure, then ask how many people will use it.

AnswerYes. Paying yearly takes 25% off, so it’s $6 per person instead of $8. How many people would be using it?

Card language “What to say?” also searches the web, even when web search is off above Voice button on cards: reads the reply aloud to you through your own speakers or headphones. Nothing is sent into the call. Voice in your ear: the card's move (its title) is spoken automatically as soon as the card is ready. Use headphones; your microphone is muted for Myna while it speaks. Sound cue: a quiet chime when a card carries a fact from your base or answers a direct question, so you look at the panel only then.

Speech model (OpenRouter) Voice (must belong to the model)

Appearance

Theme

Accent color

Custom… Applies right away. To match your Chrome theme, press “Pick from screen” and click the tab strip.

Panel font size, px

Advanced: models, speech recognition, archive, checks, API keys

Models

The quick–deep slider in the panel picks one of four model pairs. Automatic cards run on nearly every question, so their price adds up; “What to say?” runs rarely and can use a stronger model. Prices are per million tokens. “Per hour” is an upper estimate for a busy call.

Current slider position

Browse and compare models loading…

Model	In / Out per M	Context	≈ per hour	Use for

Speech recognition

Call language

Languages to expect when detecting automatically

Deepgram model

End-of-turn confidence (0.5–0.9) End the turn after a pause of, ms

Early end-of-turn signal (0.3–0.9, 0 = off). Starts the card about 300 ms earlier and cancels it if the person keeps talking. Stop listening after this many minutes without any sound (0 = never). Covers calls that end while the tab stays open. Key terms, one per line

Bridge phrases

While a card is loading, show a short phrase you can say At the start of a call, suggest how to open it (the other person has not spoken yet) Start the card while the other person is still speaking, and refine it when they finish First glance: a light model shows one or two keywords in about half a second, then the full card replaces it Your phrases, one per line (used first)

Checks

Live checks of the services Myna relies on, right from here. Results also go into the diagnostic log (the Log button in the panel).

API keys

Myna access code

With an access code, requests go through the Myna server and the keys below are not used. Leave it empty to use your own keys.

Deepgram OpenRouter

Deepgram: console.deepgram.com → API Keys. OpenRouter: openrouter.ai/keys.

Role

Knowledge

Cards

Appearance

Models

Speech recognition

Bridge phrases

Archive

Checks

API keys