google-deepmind/concordia · project editor & CLI

Build role-playing simulations with Concordia

Learn Concordia by building a fantasy world: entities, components, prefabs and engines; the project editor and the concordia-session command line; debugging with No Language Model, which makes no API calls; then live games on a real model with the sequential, simultaneous and asynchronous engines. 13 min 43 s in 14 chapters.

Chapters

    Commands shown

    Launch without a language model

    python -m pip install -e . python -m examples.project_editor.run --port 8080 --model-backend none

    Debug a prompt

    log context --source current Charlie --step 3

    Connect a real model

    python -m pip install -e ".[together]" export TOGETHER_API_KEY="your-together-api-key" python -m examples.project_editor.run --port 8080 \ --model-backend together_ai --model-name deepseek-ai/DeepSeek-V4.1-Flash

    Attach a terminal

    concordia-session --url http://127.0.0.1:8080 interactive --draft royal-cake-draft.json

    Build the game

    catalog prefabs add instance game_master.situated_in_time_and_place.GameMaster --id narrator add instance basic --id fizzwick add instance minimal --id gerald set simulation premise '"…"' validate · save · export royal-cake.json

    Inspect what you can set

    params narrator params narrator generative_clock get narrator generative_clock time

    Read compressed logs

    concordia-log actions royal-cake-log.json Gerald concordia-log dump royal-cake-log.json > inflated.json # grep on the raw log undercounts: repeated text is stored once

    Run it again

    python -m examples.project_editor.run --project royal-cake.json \ --model-backend together_ai --model-name deepseek-ai/DeepSeek-V4.1-Flash # add --headless to run without the editor

    Other engines

    python -m examples.project_editor.run --engine simultaneous \ --project royal-pond.json --model-backend together_ai --model-name … python -m examples.project_editor.run --engine asynchronous \ --project crumbleton-forum.json --model-backend together_ai --model-name … viewer Forum # or choose Forum in the central viewer

    Get the project files

    The three games from the video. Save one under the name shown and pass it with --project; the pond needs --engine simultaneous and the forum --engine asynchronous.

    Download
    
  
    

  
    Recorded on upstream main (80d9dab) plus the proposed params/get commands and editable game-master state, with the project editor example and Together AI adapter from the concordia-claw feat/android-project-editor-demo-001 branch. Chapters 3–5 use No Language Model. The live games use DeepSeek V4.1 Flash on Together AI: the 10-step cake game, the 3-round pond and the 3-round forum cost about $0.28 together. Waiting during live runs is sped up, as marked on screen. Terminal scenes replay real captured concordia-session output; long output is trimmed and marked. Narration voice: Matilda by ElevenLabs (Flash v2.5).