Configuration

Config lives in two files. Project wins.

Full example

{
  "defaultModel": "anthropic/claude-sonnet-4-5",
  "thinking": { "mode": "adaptive" },
  "genome": true,
  "semanticSummaries": "ast",
  "diffStyle": "default",
  "chatStyle": "accent",
  "toolTimeout": 2,
  "compaction": {
    "strategy": "v2",
    "triggerThreshold": 0.7,
    "keepRecent": 4
  },
  "taskRouter": {
    "spark":      "anthropic/claude-haiku-4-5",
    "ember":      "anthropic/claude-sonnet-4-5",
    "compact":    "google/gemini-2.5-flash"
  },
  "agentFeatures": {
    "desloppify": true,
    "tierRouting": true,
    "dispatchCache": true,
    "targetFileValidation": true
  },
  "retry": { "maxAttempts": 5, "baseDelayMs": 3000 }
}

Fields

Model

Display

Intelligence

Compaction

See compaction for what these do.

"compaction": {
  "strategy": "v2",
  "triggerThreshold": 0.7,
  "resetThreshold": 0.4,
  "keepRecent": 4,
  "maxToolResults": 30,
  "llmExtraction": true
}

Retry

For 429s and transient errors. Delays double each attempt.

"retry": { "maxAttempts": 5, "baseDelayMs": 3000 }

Web access

With nativeSearch on, a supported model searches in one step and returns its own citations instead of Empryo scraping on your machine — and the query goes only to your model provider rather than also to DuckDuckGo, Brave or Jina. Two trade-offs worth knowing:

• Because the provider executes the search, there is no per-query approval prompt to intercept.

Turn nativeSearch off to keep every search client-side and promptable.

• The provider's search tool adds a fixed ~2.2K input tokens to each request whether or not the

model searches. It sits in the cached prefix, so the effective cost is much lower than that.

It is on only where we have actually run it: Claude (direct and through the proxy), GPT through the proxy, and OpenRouter's native-search generations. Lanes we have implemented but never exercised against a real key stay off and use the search below instead — being wrong about a provider-executed tool costs the whole request, not just the search, so an unproven lane is not worth the gamble. /web-search names whichever backend is actually in play.

Everything else — gateways, local models, unproven lanes, no provider support — falls back to the search backends you've keyed, with free DuckDuckGo as the floor. Toggle both in Settings › Provider, or with /provider websearch and /provider nativesearch.

Task router

Assign different models to different jobs. See the task router recipe for tuning tips.

"taskRouter": {
  "spark":      "anthropic/claude-haiku-4-5",
  "ember":      "anthropic/claude-sonnet-4-5",
  "webSearch":  "anthropic/claude-haiku-4-5",
  "desloppify": "anthropic/claude-haiku-4-5",
  "verify":     "anthropic/claude-haiku-4-5",
  "compact":    "google/gemini-2.5-flash",
  "semantic":   "anthropic/claude-haiku-4-5",
  "default":    null
}

Resolution: taskRouter[slot] → taskRouter.default → active model.

Agent features

"agentFeatures": {
  "marionette": false,
  "desloppify": true,
  "tierRouting": true,
  "dispatchCache": true,
  "targetFileValidation": true
}

Toggle with /agent-features (or /marionette for the prompt compiler directly).

Marionette lane

How the prompt compiler surveys the repository before it writes the map.

"marionetteMode": "fast"

Measured on 11 real requests against a large TypeScript monorepo (haiku-4.5): fast found the right files as often as none while roughly doubling precision, and beat deep by 30 points of recall at a third of the latency. deep keeps the highest ceiling on requests whose product wording matches nothing in the code, at the cost of high run-to-run variance.

Switch with /marionette none|fast|assist|deep, in Settings → Router → Behaviors, or per run with --marionette-mode <lane>. marionetteBudget (steps / time) only applies to assist and deep — the retrieval lanes have no loop to bound.

Providers

Pick keys and add custom providers on the dedicated pages:

• All providers - env vars, --set-key commands, keyUrls.
• Custom providers - any OpenAI-compatible API.
• Copilot - use your GitHub Copilot subscription.

Auth & key storage

Keys go to your OS keychain when available, with a file fallback otherwise.

empryo --set-key <provider> <key> writes to whichever backend is active. Inspect with empryo --list-providers.

Lookup order

{ "keyPriority": "env" }

Use "app" when shell-exported keys keep overriding stored ones.

Pass a key per-launch

Env vars work for the TUI too - useful for one-off sessions or CI:

LLM_GATEWAY_API_KEY=sk-... empryo
ANTHROPIC_API_KEY=sk-ant-... empryo

Nothing is persisted. There is no --key flag - the TUI only reads stored keys and env vars.

Remove a key

/keys inside the TUI lists every provider with delete shortcuts. Or delete from the keychain directly (security delete-generic-password -a empryo -s anthropic-api-key on macOS).

Instruction files

Auto-load project rules from AI-tool markdown files:

{ "instructionFiles": ["empryo", "claude", "cursorrules"] }

Toggle in TUI with /instructions.

Privacy

Block files from the agent:

/privacy add .env
/privacy add secrets/**

Built-ins already cover .env, .pem, credentials, id_rsa, .npmrc, .netrc, shadow, passwd.

Environment variables

Hooks

Covered on the hooks page. Short version: Empryo reads hooks.{Event} from all 5 config sources (Claude Code's three + Empryo's two), fires all matches.

Storage

/storage shows per-component disk usage (Genome, sessions, plans, memory, input history, binaries, fonts) with one-click cleanup.

Scope priority

Session → Project → Global. Changes via commands land in whatever scope the command targets. Use /model-scope to move the active model between project and global.