Skip to content
Attest Framework

Configuration Reference

All Attest configuration is driven by environment variables. No config files are required.

Engine Discovery

Section titled “Engine Discovery”
VariablePurposeDefault
ATTEST_ENGINE_PATHAbsolute path to engine binary — skips all discoveryunset
ATTEST_ENGINE_NO_DOWNLOAD1 / true / yes disables auto-downloadunset (download enabled)

The SDK resolves the engine binary in this order:

ATTEST_ENGINE_PATH
  → PATH lookup
  → ~/.attest/bin/ (shared cache, version-checked)
  → ../../bin/ (monorepo dev layout)
  → ./bin/ (local)
  → auto-download from GitHub Releases
  → error

Engine Runtime

Section titled “Engine Runtime”
VariablePurposeDefaultSince
ATTEST_ENGINE_TIMEOUTMax seconds to wait for an engine response before the SDK raises a timeout error30v0.5.1
ATTEST_SIMULATIONEnable simulation mode — deterministic mock results, no engine process, no API callsunsetv0.4.0

Cost & Budget

Section titled “Cost & Budget”
VariablePurposeDefaultSince
ATTEST_BUDGET_MAX_COSTMaximum USD spend per evaluation. Assertions that would exceed this limit are skipped with a budget error. Applies to judge (L6) and embedding (L5) assertions.unset (unlimited)v0.5.1
ATTEST_RATE_LIMIT_OPENAIRate limit for OpenAI API calls (requests per minute)unsetv0.4.0

Judge Cache

Section titled “Judge Cache”
VariablePurposeDefaultSince
ATTEST_JUDGE_CACHE_MAX_MBMaximum size in MB for the LLM judge response cache. Uses LRU eviction. Set to 0 to disable caching.100v0.5.1
ATTEST_CACHE_DIRDirectory for all Attest caches (judge, embeddings, engine binary)~/.attest/cachev0.4.0

History Store

Section titled “History Store”
VariablePurposeDefaultSince
ATTEST_HISTORY_MAX_ROWSMaximum number of rows in the SQLite history store. Oldest rows are deleted when the limit is exceeded.10000v0.5.1
ATTEST_HISTORY_MAX_AGE_DAYSAuto-delete evaluation results older than this many days90v0.5.1

Continuous Evaluation

Section titled “Continuous Evaluation”
VariablePurposeDefaultSince
ATTEST_CONTINUOUS_QUEUE_SIZEMaximum number of pending evaluations in the continuous eval queue. When full, new submissions block until space is available (backpressure).1000v0.5.1

LLM Providers

Section titled “LLM Providers”
VariablePurposeDefault
OPENAI_API_KEYOpenAI API key for judge and embedding assertionsunset
ANTHROPIC_API_KEYAnthropic API key for judge assertionsunset
ATTEST_GEMINI_API_KEYGoogle Gemini API keyunset
ATTEST_OLLAMA_BASE_URLOllama server URLhttp://localhost:11434

Judge Configuration

Section titled “Judge Configuration”
VariablePurposeDefault
ATTEST_JUDGE_PROVIDERLLM provider for judge assertions: openai, anthropic, gemini, ollamaopenai
ATTEST_JUDGE_MODELModel name for judge assertionsgpt-4.1

Embedding Configuration

Section titled “Embedding Configuration”
VariablePurposeDefault
ATTEST_EMBEDDING_PROVIDEREmbedding provider: auto, openai, onnxauto

auto selects ONNX (free, local) when available, falling back to OpenAI.

Example .env

Section titled “Example .env”
Terminal window
# Engine
ATTEST_ENGINE_TIMEOUT=30
ATTEST_SIMULATION=0


# Budget
ATTEST_BUDGET_MAX_COST=1.00


# Cache & history
ATTEST_JUDGE_CACHE_MAX_MB=100
ATTEST_HISTORY_MAX_ROWS=10000
ATTEST_HISTORY_MAX_AGE_DAYS=90
ATTEST_CONTINUOUS_QUEUE_SIZE=1000


# Provider keys
OPENAI_API_KEY=sk-...
ATTEST_JUDGE_PROVIDER=openai
ATTEST_JUDGE_MODEL=gpt-4.1