Configuration

All configuration lives on the Soul node on the graph — no files. Change it with configure <key> <value> (CLI or the cfg_* chat tools), or interactively with sigil config. Every change persists on the graph and takes effect on the next turn.

Current values are shown by sigil soul (or /soul in chat).

Keys

Key	Values	What it does
`name`	text	The agent's name.
`persona`	text	Voice/character the agent speaks with.
`chat_model`	model name	The user-facing model for chat and sub-agents — use a strong one. Falls back to `frontier_model`.
`frontier_model`	model name	Expensive model that compiles new skills.
`small_model`	model name	Cheap model that executes compiled skills.
`router_model`	model name	Model that routes a task to a skill. Falls back to `small_model`.
`workspace`	path	Default directory the tools work in — where relative paths resolve and shell jobs start. Default `~/.sigil/workspace`. See workspace-and-tools.
`sandbox`	`off` \| `jail` \| `docker`	Containment for the file/shell tools. Default `off` — full host access, so Sigil can drive the tools and agents already installed here. `jail` confines the file tools to the workspace; `docker` also runs each command in a locked-down container.
`verbose`	on/off	Show the model's output as it is generated. Default on. A compile is dozens of model calls and the words being written are the artifact; `--quiet` or `SIGIL_VERBOSE=0` turns it off for one run. See skill-compilation.
`prompt_mode`	`full` \| `minimal` \| `none`	How much context (ethos + memory) wraps the compiler/run.
`auto_eval`	on/off	Grounded-eval valve: judge each skill run and relearn if it fails.
`eval_threshold`	0–100	A run scoring below this auto-triggers a relearn + retry.
`recall_mode`	`lexical` \| `vector` \| `hybrid`	How semantic memory is retrieved. See memory-and-skills.
`embed_model`	model name	Embedding model (litellm name) for vector/hybrid recall.
`add_ethos`	text	Append an operating principle to the ethos.
`add_channel`	name	Register a channel name on the soul.
`tool_allow`	csv	If non-empty, only these tools (+groups/profile) are permitted.
`tool_deny`	csv	Denied tools (deny wins over allow); `*` denies all.
`tool_profile`	group name	Active profile — a tool group whose tools join the allow-list.

Compiled skills are tools too, named skill_<signature> — so tool_deny skill_clean_csv takes that skill off the agent's belt and out of what sigil mcp-serve publishes. sigil tools skills shows the current list and the gate in front of each; see chat-and-tools.

Examples

sigil configure chat_model gpt-4o-mini
sigil configure small_model ollama_chat/qwen3:8b
sigil configure workspace ~/projects/sigil-ws
sigil configure sandbox jail          # opt back into containment
sigil configure verbose off           # stop streaming model output
sigil config                       # interactive editor for the common keys

In chat, the agent can change its own model, persona, or prompt mode when asked (via the cfg_set_model / cfg_set_persona / cfg_set_prompt_mode tools).

Model tiers

chat (talk to it) · frontier (build skills) · small (run skills) · router (route tasks). Manage names, aliases, and fallback chains with sigil models … — see models.