Panel reference

This page takes the client panel apart feature by feature. If you just want to get playing, Install and first run is enough.

The header

Element	What it does
Connection status	Whether you’re connected
Vitals	Your character’s live numbers
Spectator memory use	How much memory the 3D view is holding
Reload	Reloads the 3D spectator view (use it if the view freezes or corrupts)
Pause / Resume (⏸ / ▶)	Pauses a running agent. While paused it makes no LLM calls, so it costs nothing — use it to take over by hand, or to save money
Apply & restart	Only appears if you changed a setting while the agent is running. The agent reads its config once at startup, so this is how a change takes effect
Manual / AI switch	Swap between the two modes live
Change character / server	Back to the earlier screens

The six drawers

Equipment

What’s worn, slot by slot.

Why equipment comes through the relay

The agent’s own panel API reports the bag but never the gear. So this drawer is reconstructed from the inventory frames the relay sees — one of the things the desktop client adds that running agent-client directly does not give you.

Bag

Shows what’s carried, and lets you mark what the agent is allowed to dispose of:

• S (Sellable) — the agent may sell it
• D (Dropable) — the agent may drop it

Tick the boxes, then press Apply labels to send them (they’re staged, not sent one at a time).

This is the gate that stops an agent disposing of gear you care about — leave something unmarked and it won’t touch it.

Personality & Memory

Two halves:

• Personality — this character’s own prompt, layered on the shared rules. Editable. Saving while the agent is running restarts it.
• Memory — what this character remembers, written by the agent itself and read-only.

Personality is where you shape behaviour (“you are a cautious hunter; retreat below half health”). Memory is what it has accumulated on its own.

Coordinates

Place bookmarks. Use current position records where you are, then name it and add it.

Pairs with Dispatch: save fishing spots, dungeon entrances and shop locations, then send the agent to one.

Dispatch Presets

Save frequently used instructions under a name and send them with one click instead of retyping.

Each preset has a name and the prompt to send. Saved presets appear as buttons beside the Dispatch input.

Activity

Two tabs:

• Thoughts — every prompt sent and every reply, with timings. Has filters, Follow (auto-scroll to newest) and Clear.
• Log — the agent process’s own stdout / stderr.

Wondering why the agent made a decision? Thoughts. Wondering why it broke? Log.

Dispatch

The strip under the game view is the one control that reaches a running agent.

Type an instruction and it arrives as the character’s next turn — best-effort, with no guarantee it will be followed. It’s a whisper, not an order: you’re suggesting, the model still decides.

The four settings tabs

LLM

7

Field	Default	What it does
Backend	Codex CLI	One of four — see below
Model	Varies by backend	Suggestions in a dropdown; free text also works
Base URL	https://ollama.com/v1	Only appears for the OpenAI-compatible backend
API key	—	HTTP backends only. Stored in the OS keychain
Reasoning effort	none	none / low / medium / high. Only affects models that support it
Maximum tokens	4096	Cap on a single reply
Temperature	0.7	Lower is steadier, higher is more varied

The four backends

4

Backend	Kind	Suggested models	Notes
Codex CLI	CLI	gpt-5.4-mini, gpt-5.4, o4-mini	Uses the Codex CLI you are already signed in to. No API key needed here
Claude CLI	CLI	sonnet, opus, haiku	Uses your local Claude CLI login. Also no key
OpenRouter	HTTP	qwen/qwen3.7-flash and others	Needs an API key. Model guidance on the previous page
OpenAI-compatible	HTTP	Whatever you enter	Any OpenAI-compatible endpoint. Defaults to Ollama, so local models work

Already have Claude or Codex CLI? You don’t need an API key

The CLI backends reuse the login you already have on this machine.

Running fully local

Pick OpenAI-compatible, put your Ollama address in Base URL (the default is already in Ollama’s shape) and name your local model. No API cost and no data leaves the machine — the trade is that local models are often slower than qwen3.7-flash and may not keep up with the game’s turn cadence.

Agent

The old Automatic play and Advanced tabs were merged into this one in v0.18.

9

Field	Default	What it does
Active response	5s	Response cadence when something is happening. Shorter is more responsive and more expensive
Quiet-world patrol	8s	Response cadence when nothing is going on
Continue adventuring while alone	On	Turn it off and it stops calling the AI when no other players are nearby — the setting to use when idling
Active interval (s)	5	The raw value behind the slider above
Idle interval (s)	8	Likewise
Watch port	8808	Port for the local spectator panel. Change only on a conflict
Log level	info	info / debug / warn. Raise to debug before reporting a problem
Concurrent LLM calls	2	How many requests are in flight at once
Request timeout (s)	120	How long to wait for one request. Raise it for slow local models

The first three are the main cost control. To spend less, raise both intervals or turn off adventuring while alone.

Display — the look of action toasts

The agent raises a toast for each action it takes. This tab styles them, with a live preview.

5

Field	Default
Font size	13
Background transparency	75
Persist time	7s
Fade duration	0.4s
Max toasts shown	10

Purely cosmetic — none of it changes agent behaviour. If the screen feels busy, lower the count and the persist time.

Audio

2

Field	Default	What it does
BGM volume	100	Background music volume, mutable
Sound effects	100	Sound effect volume, mutable

These control the embedded game view’s own volume. The panel pushes them in via postMessage; the real storage is that iframe’s own localStorage (a different origin), so this is only what gets pushed to it.

Connection profiles

Managed on the Server screen. The built-in openmmo.to.nexus profile is fixed; custom ones can be created, edited, duplicated and deleted, and test-connected before use.

A profile holds:

Field	What it is
Profile name	Name of the profile
Game server	The server’s WebSocket address
Terrain origin	Where terrain assets are fetched from
Google client ID	The OAuth client that server accepts
Google client secret	Its secret, stored encrypted

These five travel as a set — a server only accepts sign-in tokens issued by a client ID on its own allowlist, so client IDs can’t be shared between servers.

Where your secrets live

API keys and connection-profile secrets are encrypted with the OS keychain (Electron safeStorage, with an AES-GCM fallback) and handed to the agent as environment variables.

They are never written into agent-client’s config file — a config pasted into a bug report carries no credential. The config itself is regenerated from the panel on every start; an existing hand-written file is imported once and backed up next to it.