Tablefolk — Local Agent Toybox

Preparing local brainruns locally on this device

Checking local model cache…

Adventurer

View and pace

Controls

WASD moveclick the ground: walk there E open / pick up / talkF attack V first personC follow / free camera H window viewP let them live I this panelEnter chat Esc close panel

Settings

You

Developer mode

All the panels: world, characters, scenario editor, library and debug, the brain loop and the command box.

Messages

World

Load incident state Drop a character APNG here or click to choose Restores its embedded initial setup, not the animation playback.

Spawn

Map

Base map Walls campfire my avatar

Click-to-spawn mode

enable click-to-spawn

Danger zone

clear map and reset state keep character customizations. only the button above wipes them.

Character

Agent Body Mind style Mind enabled Fireball temptation Chaos Quest Current goal

Personality / experiment note thinking mode (reasons before acting: slower, smarter) show thought text in a bubble (otherwise just "thinking…" while it reasons)

Inventory

Gold

Items, one per line (optional count: "healing potion, 3")

Character voice

Speak aloud Voice

Scenario

A scenario is a little game with a goal, shareable as a link. Playing one never touches your sandbox world; "exit" brings it back.

Scenario (JSON)

Map

Describe the place

Map lines (e.g. place barrel x3 beside counter)

seed

Test tactics

Plays the scenario above automatically, several times per tactic, and counts wins. Each "## name" block is one tactic; its lines are your messages, one per turn. "## llm: strategy" lets the model play you instead (add "turns: N"). Stop the running brains first. Tip: open the page with ?norender=1 so the model gets the GPU to itself. Tactics

runs each

Things: crate, barrel, mushroom, lamp, wall, counter, note "text", or a character id. Placements: at center/north/…/random, against north_wall, beside / near / behind / in_front_of / around X (X: anything placed earlier, a character, house, campfire), plus row / cluster / scattered. Also: path from X to Y. Goals and the full format: docs/SCENARIOS.md.

Library

Scenes (world setups)

Saved characters

saving a character snapshots the currently selected character's body, personality, voice and settings.

Packages, skills, roles, asset packs

The official list (registry/index.json), plus npm packages and GitHub repos tagged tablefolk. Community ones are not reviewed; ones marked "code" ask before their code runs.

GitHub too

Code allowed in this browser

Voice

speak say() text aloud

Engine Language Breeze server

Volume Buffering

Buffering trades startup delay for glitch resistance. Takes effect the next time speech starts fresh in this tab (reload if you've already spoken this session).

Pocket-TTS off

backend: WASM CPU · waiting for speech metrics

Speech input (the 🎤 button)

Custom voice sample

Use roughly 5–15 seconds of clean speech. Samples stay in this browser.

Graphics

window view (H): the webcam follows your head and the screen becomes a window onto the board leaning in Face tracking runs in this browser (MediaPipe, about 14 MB once); the video never leaves the page, and a small preview shows while it is on.

Everything below defaults to the original look. Turning things down trades a little visual fidelity for smoother frames - worth trying if the world feels heavy or you notice audio glitching while things render.

Quality

Scene

Light (the default) has no shadow maps or post effects, leaving the graphics card to the characters and their voices. Medium adds detail with 1k textures and no ambient occlusion; high is full detail. Every setting uses the simple sea. Changing this reloads the page.

While a character is thinking

Render While Breeze speaks

The model shares the GPU with the 3D view. Lower = faster replies; paused freezes the view until the reply arrives.

Shadows

key light shadows (island) Refresh every frame(s) campfire shadows Refresh every frame(s)

Each is a real-time point-light shadow (renders 6 cubemap faces every refresh). "Refresh every 2 frames" halves that cost; disabling stops the map updating entirely.

Sea

wave animation (vertex displacement) shoreline foam ribbons

Both rebuild geometry on the CPU every frame. Foam is the cheaper/less-noticeable of the two if you just want to shave a little cost.

Characters' memory

smart memory (EmbeddingGemma 2, 175 MB once)

Characters recall what they know by meaning, not only by matching words ("where should I look?" finds "the wolf den is in the hollow"). Runs in the browser; off: keyword matching only.

Debug

ready

Brain options

OpenAI-compatible endpoint (for the "endpoint" brain)

Preset

Base URL API key Model Server's models

vision: send each character a picture of what it sees (needs a model with image input: gemma4-e2b-mm on litert-lm serve, Gemma 4 / GPT on OpenRouter; slower)

Any OpenAI-compatible server: this Tablefolk server's proxy (default), OpenRouter, OpenAI, litert-lm serve, llama.cpp, Ollama (a server on your own computer must allow this page's requests: CORS). The key stays in this browser; an OpenRouter preset uses the key you signed in with on the world builder.

chat test mode (bare conversation, no perception) anti-echo instructions (off by default - small models can confuse negative prompts)

Cache

Perception & prompt

Vision radius ASCII map

Load brain and step to see prompt...

Performance

live updates

perf monitor idle - check "live updates" to sample

History

to: Chat name Your gold WASD / click: move · E: open, pick up, talk · F: attack · I: inventory · V: first person · C: free / follow camera · P: let them live

Pick a character (Character panel) to shape its mind, chat below, then start the local brain from the top bar.

This browser cannot run Tablefolk yet.