AI Agent (Eden)¶

EVE Console ships with an optional built-in conversational assistant — Eden by default (you can rename it). It has read access to your local database and a set of tools, so it can both answer questions the UI isn't explicitly designed for and do things in the app for you.

The agent is entirely optional and inactive until you set it up. If you don't configure it, you can ignore it completely. Its settings live in Settings (the ⚙ gear button, top-right) ▸ AI Agent, across the Agent, Personalisation, Voice (TTS) and Speech Input (STT) tabs.

What it can do¶

Answer questions about your data. It discovers your local database's shape and runs read-only queries itself, so it can pull together things no single screen shows — assets, industry jobs, market prices, character info, and more.
Reach ESI when the database doesn't have it. For live details the app hasn't synced, it can make a read-only ESI call.
Act in the app for you. Open any tool, navigate to an item or entity, open the map with an overlay, set a destination, filter the asset or industry views, configure the Item Browser, select a character, refresh data, and create or manage alarms.
Know what you're looking at. "This", "here" and "this screen" mean the tab you have open, so you can ask "what does this show?" without naming it.
Show its work in its own tab. Results can be rendered as a table or a document you can keep, rather than only as chat text.
Know who you are. It follows your name and standing instructions (see Personalisation) so its answers fit how you play.
Run on the models you choose — a free model on a server of your own (Ollama, LM Studio) or a paid service (Claude, OpenAI), or several, each doing the job it suits (see Roles).
Talk and listen — optional text-to-speech and speech-to-text with a global push-to-talk key, so you can use it hands-free while the game has focus.

Once enabled, a ✦ {name} button appears in the title bar; click it to toggle the (resizable) chat panel. The line above each reply shows when it was written, which model wrote it and which tools it used.

Setup¶

Agent settings are split between tabs on purpose:

The Personalisation tab is about you, so it's kept in the database and shared by every client that opens your data.
The Agent, Voice (TTS) and Speech Input (STT) tabs are this machine's own (models, API keys, voices, microphone).

Everything on these tabs saves as you change it — text once you pause typing or leave the box. The line at the bottom confirms each save.

Defaults keep up with the app

Texts you can reword — the router's rule and the announcements below — are stored only when you change them. Left as they are, they follow any improved wording in later versions.

Agent tab¶

Enable — tick Enable {name} AI companion. This just makes the panel available; you still need at least one model set up below.
Models — every model the agent can think with. Click Add for a new one and select it in the list to edit it; Remove takes the selected one out (at least one model stays). For each model:
- Name — optional; what the roles and announcements call it. Blank shows what it is, e.g. Local — qwen3:8b on gpu-box.
- Service — Local server — Ollama, LM Studio… (free), Claude — Anthropic (paid) or OpenAI (paid).
- The service's key or address:
  - Claude (Anthropic) API key (sk-ant-…) — one key for every Claude model in the list. Get one at console.anthropic.com.
  - OpenAI API key (sk-…) — one key for every OpenAI model, and also used for OpenAI's voices and speech input. Get one at platform.openai.com.
  - Server address for a local server, e.g. http://gpu-box:11434 (the server's root). It must offer OpenAI's /v1/chat/completions, as Ollama, LM Studio and most local runners do.
- Model — chosen from the service's own list: the models your Claude or OpenAI key can use, newest first (OpenAI's list leaves out models that aren't for conversation), or the models on your local server. Click ↺ to ask again. A model you chose earlier that is no longer listed stays chosen, marked — not in the list now. Without a key or an answer from the server, the reason is shown and you can type the name instead. There is no default model: one with none chosen isn't set up.
- Let it think before answering (slower) — local servers only, on by default. For models that think before they answer, such as Qwen3; the thinking is never shown or spoken. Untick it for quicker replies at some cost in care. Models that don't think ignore it.
- Prompt cache lifetime — Claude only: 5 minutes (default) or 1 hour. Reading the cache costs a tenth of sending the conversation again; the 1-hour option saves money if your messages are usually 5–60 minutes apart, but writing it costs more, so it costs more if they are closer together. The AI Usage & Cost tool shows cache reads and writes per message. Applies to every Claude model in the list.
- Test This Model — asks it for one sentence, as set on the tab. It shows how long it took and what it said, or why it failed. Free on a local server; a few tokens on a paid service.
Roles — which model does what (see Roles below).
When a model changes — what is said when a role falls over to its fallback and when its own model comes back (see Fallbacks).
Context Management — optionally persist chat history to disk across restarts (Persist conversation history across sessions) and set a Summarization threshold (estimated tokens) at which older messages are silently compacted into a summary. Lower values cut cost per message but drop older context.

Older settings

Settings from before models and voices became lists are brought across on first start: your one model becomes the first in the list and does every job, and your one voice becomes the first voice. Nothing changes until you change it.

Roles¶

The Roles section names a model for each job. The lists show each model with (free) or (paid) after it.

Conversation — talks with you and works the app: opens and arranges its tools, sets alarms and destinations, remembers your standing instructions.
Questions about your data — anything that needs your database or ESI: what you have, own, are doing or have done. This model gets the whole prompt and every tool. Leave it on Same as the conversation for one model to do everything.
Summaries — compacts older messages when the conversation grows long. A small free model does this well. A local summary model gets a short one-line prompt rather than the agent's full one, leaving its window for the history being summarised, and the history is checked against a local model's window after every turn so it can't silently outgrow it.

The sentence under the section sums up what your choices come to.

When Questions about your data names a different model from the conversation:

Before each message, the conversation model is asked in one word whether the message needs your data. If it does, the data model answers it instead, with the database and ESI. The conversation model gets a prompt about half the size and no access to your data.
Send a message to this model when… holds the rule the conversation model decides by. Reword it if messages go the wrong way; Reset to the default puts the app's wording back.
If a data question slips through, the conversation model hands it over itself.
Start a message with /data or /chat to send that one message to the data model or the conversation model regardless.

Fallbacks¶

The Conversation role — and Questions about your data, when it has a model of its own — can name a model to use when its own stops answering (If it stops answering, fall over to). None — the role simply fails turns this off.

A model that stops answering is replaced at once, before a word of the answer.
Whether a model is there is checked for free (the service's own model list), never by paying for an answer.
Nothing is switched silently at start: if a role's own model isn't there, the first message finds out and says so.
Say so when it falls over, and when it comes back — in the chat, and aloud when speaking announces each change.

Under When a model changes, Falling over and Coming back are the texts said. {user} is your name from Personalisation, {purpose} is "our conversation" or "data access", and {primary} and {fallback} are the two models' names. The two minute settings — At least … minutes between one change and the next switch back and Up for … minutes without a break before a role's own model returns — only govern coming back, so a server that keeps dropping out can't make the agent flip back and forth. A switch back also waits for your next message.

Personalisation¶

On the Personalisation tab (shared across your clients):

Agent Name — blank = Eden.
Response Verbosity — Concise (1–3 sentences), Balanced, or Detailed. This is added to the system prompt on every message.
Your Name — how the agent addresses you.
Standing Instructions — free-form, lasting guidance for the agent: how you play, what you care about, conventions to follow. It's included on every message, so keep it focused. The agent can also update this itself when you ask it to remember something.

AI Usage & Cost¶

The AI Usage & Cost tool records what each turn cost and did — tokens in and out, cache reads and writes, and the estimated price per message — so you can see where spend goes and whether the prompt-cache setting is paying off. Local models are free to run.

Its Rates tab holds the price of each paid model. Every model a paid service lists for your key gets a row of its own. Where LiteLLM's community price list has the model, the row is filled from it (the list is fetched again once a day), and the row's note names the list, the date and the provider's pricing page to check it against. A rate you set by hand is never overwritten. ElevenLabs voices aren't in that list, so set their rates yourself.

Voice (optional)¶

On the Voice (TTS) tab, tick Speak the agent's replies and build a list of Voices, in order of preference. The first voice that can speak is used; if it fails, the next takes over at once, and the first returns once it's back and has stayed up. Use Add, Remove, ▲ Up and ▼ Down to arrange the list.

For each voice:

Name — optional. The voice is part of who you're talking to: while a named voice speaks, its name is the agent's name, in the panel and in the conversation. A voice with no name uses the agent's name from Personalisation.
Engine:
- Kokoro (on this PC, free) — high-quality neural TTS that runs fully offline. Pick a voice, then click Load Model: one model serves every Kokoro voice and is downloaded once (~320 MB of neural-network weights, no executable code) into this PC's data folder.
- Local server — Chatterbox, Orpheus, Kokoro-FastAPI… (free) — any server that speaks OpenAI's speech API, usually on another machine's GPU. Set the Server address (up to and including /v1), e.g. http://gpu-box:8880/v1, the Model and Voice as the server names them, the Speed, and an API key only if the server asks for one. A server counts as up only if it actually speaks within 30 seconds.
- Piper (on this PC, free) — fast offline neural TTS; the runtime is bundled, and only the chosen voice (a data file) downloads, with Download Voice Model.
- OpenAI (paid) — OpenAI's voices, billed per character. Uses the OpenAI API key (the same one as the Agent tab's OpenAI models). Choose the Model from OpenAI's list (↺ asks again) and a Voice from the ones OpenAI documents for that model, and set the Speed.
- ElevenLabs (paid) — ElevenLabs' voices, billed per character. Paste your ElevenLabs API key, then choose a Voice by name from your account and a Model from ElevenLabs' list (↺ asks again, e.g. after adding a voice there).
Test This Voice — speaks with this voice as set, on its own rather than through the list. It shows how long it took in green, or the reason it couldn't speak (the server's own words) in red.

All voices are brought to the same loudness, so a change of voice doesn't jump in volume. Speech leaves out emoji and markdown — headings, rules and table lines aren't read out — and the next sentence is prepared while one plays. Volume and mute are also available in the agent panel itself.

When the voice changes — tick Announce it, in the new voice to have the voice taking over say so. Taking over and Coming back are the texts, with {previous} and {current} for the two names. A change is announced only when the name changes, so give each voice its own name. Nothing is said when the app starts. The two minute settings work as they do for models: a failing voice is replaced at once, and a switch back waits for the end of the current answer.

Speech Input (optional)¶

On the Speech Input (STT) tab, pick a Provider to talk to the agent:

OpenAI Whisper — cloud transcription (shares the OpenAI key); Windows and Linux. Choose the Model from OpenAI's list of transcription models (↺ asks again). Settings from before this choice existed keep whisper-1.
Local Whisper — runs on this machine, no API key; Windows only. Pick a Model and click Download Model before first use, and set the Language you speak as a two-letter code such as en (or auto to let the model guess, which is often wrong on a short clip).

Then choose a Microphone Device (click ↺ to refresh the list) — leave it on System default to follow whatever Windows/your OS is set to — and a Global Push-to-Talk Key — hold it to record even when EVE has focus. F13–F20 are rarely captured by games and make good PTT keys. The mic button in the panel works regardless of this setting.