Environment variables

An app reads its settings from .env, at the root of the app, which is never committed: it holds what changes from one machine to another and every secret. mfw create writes .env.example with every variable the chosen pieces read; copy it to .env and fill it in. After changing it, mfw start applies the change.

These are the core’s. Each plugin’s are on its page.

The model

Variable
OLLAMA_HOST	Required. The Ollama server. Ollama on the machine running Docker is http://host.docker.internal:11434.
OLLAMA_MODEL	Required. The chat model. It must support tool calling.
OLLAMA_EMBEDDING_MODEL	The model that turns text into numbers for the long-term memory (default nomic-embed-text).
OLLAMA_THINK	false for a model that doesn’t support thinking (reasoning before it answers).
TURN_STALL_TIMEOUT_MS	How long an answer on a streaming channel (HTTP) waits for the model’s next word before it stops with an error (default 120000, two minutes; 0 turns it off). Time spent in a tool doesn’t count.

Memory

Variable
WIKI_VAULT_PATH	Required. Where the wiki lives: in a scaffolded app, the Docker volume mounted at /app/wiki-vault.
QDRANT_URL	Qdrant, for the long-term memory (default http://qdrant:6333, the service the app’s compose file starts).
SESSION_IDLE_TIMEOUT_MS	How long a quiet conversation keeps its history in memory (default 30 minutes).
SESSION_RESUME_MAX_CHARS	How many characters of a reopened conversation come back to the model (default 30000; 0 turns resuming off). More gives the model more of the past, and makes every answer slower.

Memory explains what each of these is for.