Modes, rooms & routing

How Vanta adapts its stance, scopes itself per project, routes work across models, and improves itself over time.

Operator modes

Modes are real skills that set Vanta's working stance. Install and run them:

vanta modes install
vanta skill <mode> "<instruction>"

Mode	Stance
silent-executor	Do the work, minimal narration
collaborator	Think alongside you, surface tradeoffs
critic	Adversarial review of a plan or output
researcher	Gather + synthesize before acting
debugger	Systematic root-cause investigation
assistant	General help
solutioning-mode	Goals + research → a ranked what-to-build recommendation, then stop before implementing

Auto stance — mode-detect infers the right stance from your message and prepends a hint to the turn (disable with VANTA_MODE_DETECT=0).

Project rooms

A room runs Vanta rooted in a specific project, with its own goal stream:

vanta rooms                      # list projects under VANTA_PROJECTS_DIR
vanta room <name> "<instruction>"  # run rooted in that project

VANTA_PROJECTS_DIR defaults to ~/Documents/GitHub/_active. Each room keeps its own .vanta/ (goals, events, approvals), so work in one project never bleeds into another.

Model routing

Run cheap tasks on a small/local model and hard tasks on a strong one — automatically:

VANTA_MODEL_CHEAP=ollama:qwen2.5:14b
VANTA_MODEL_EXPENSIVE=openai:gpt-5.5

Vanta classifies each task (cheap vs expensive) and routes it. Unset = no routing (everything uses the active provider). Vision is routed separately — see Providers.

Self-improvement

After turns, an opt-in background pass reviews the transcript and captures durable, reusable skills (tagged as learned) into ~/.vanta/skills, and distils 0–3 durable memories into the brain. A usage tracker proposes capturing a workflow as a skill once it recurs.

VANTA_SELF_IMPROVE=1     # capture skills from successful turns
VANTA_BRAIN_LEARN=1      # distil memories post-turn

Best-effort and gated — see Skills & memory.