blumi reads layered configuration (via figment):

1. Built-in defaults
2. ~/.blumi/settings.json (global)
3. ./.blumi/ in the current project (per-project overrides)
4. Environment variables, prefixed BLUMI_ with __ as the nesting separator (e.g. BLUMI_GRID__SECRET, BLUMI_LLM__MODEL)

Secrets live in settings.json, which is written with mode 0600. Never commit it — the repo's .blumi/ is gitignored.

blumi login

Pick a provider, paste a key (or endpoint), and choose a model. It writes the right bits into settings.json. You can re-run it any time to switch providers.

blumi is provider-agnostic. Supported providers include:

Override per-invocation without editing config:

blumi --provider anthropic --model claude-sonnet-4 run "..."
blumi --provider openai tui

A minimal settings.json:

{
 "llm": { "provider": "anthropic", "model": "claude-sonnet-4" },
 "providers": {
 "anthropic": { "api_key": "sk-ant-..." },
 "openai": { "api_key": "sk-...", "base_url": "https://api.openai.com/v1" }
 }
}

settings.json also holds (all optional, sensible defaults):

• permissions — which tools auto-allow / ask / deny. Interactive by default; toggle YOLO (auto-approve everything) with ctrl+y / /yolo / --yolo.
• persona + personas — the active agent persona and your custom ones (e.g. architect, pair, reviewer). Switch with --persona <name> or /persona.
• executor — local (host) or docker (sandbox); ssh for a remote host.
• web — gateway password hash (set via blumi serve pair).
• voice — see Voice.
• grid — see Grid.
• brain — local-LLM approval reviewer (off / advisory / auto).
• acceleration — GPU/accelerator mode for the bundled embedder (auto / cpu / apple / cuda). See Memory & Knowledge → GPU acceleration.
• heal — self-healing & evolution (reflex recovery budget, failure→fix learning, autonomy). See Self-Management → Self-healing.
• embeddings / memory / knowledge — local embeddings, semantic memory, and the code knowledge base. See Memory & Knowledge.
• mcp_servers, lsp_servers, gateway (messaging bots), git (commit identity).

When a turn stops only because it hit the per-turn tool cap, blumi auto-continues the same session and narrates each step. It's bounded by both a step budget (llm.max_auto_continue, default 12) and a token ceiling (llm.max_auto_continue_tokens, default 400k) — whichever is hit first. Tune live in the TUI with /autocontinue <n> (0 disables).

Two markdown files the agent reads/writes: project MEMORY.md and your USER.md (in ~/.blumi/). View/edit them in the TUI (/memory), the web/phone control center, or on disk.