Runtime Configuration
Agnitra uses environment variables to keep the CLI, SDK, and control plane stateless. This reference lists the most important settings and when to apply them.Core Credentials
| Variable | Description |
|---|---|
AGNITRA_API_KEY | Required credential for CLI, SDK, and API calls. Scope keys per environment and rotate every 90 days. |
AGNITRA_API_BASE_URL | Overrides the managed control plane hostname (e.g. https://api.agnitra.ai). Use it for staging or self-hosted deployments. |
AGNITRA_CONTROL_PLANE_URL | Internal service discovery override when deploying the Starlette control plane behind a load balancer. |
Telemetry & Usage
| Variable | Description |
|---|---|
AGNITRA_TELEMETRY_EXPORT | Comma-separated list of telemetry exporters (s3,stripe,aws,internal). |
AGNITRA_DASHBOARD_BASE_URL | Custom dashboard endpoint for multi-tenant deployments. |
AGNITRA_USAGE_WEBHOOK | Optional HTTP endpoint that receives every usage event for auditing or enrichment. |
AGNITRA_USAGE_SIGNING_KEY | Secret used to sign outbound usage events. Rotate regularly if you rely on downstream verification. |
Licensing & Compliance
| Variable | Description |
|---|---|
AGNITRA_LICENSE_ORG / AGNITRA_LICENSE_SEAT | Explicit license metadata for on-prem installs and offline validation. |
AGNITRA_REQUIRE_LICENSE | When set to 1, optimization runs abort if license validation fails. |
AGNITRA_OFFLINE_MODE | Forces the CLI/SDK to skip control-plane calls (also available via --offline). |
Optimization Behaviour
| Variable | Description |
|---|---|
AGNITRA_ENABLE_LLM_RL | Enable reinforcement-learning passes after Responses API suggestions. |
AGNITRA_ONLY_LLM | Skip PPO and apply only LLM-suggested kernels (useful for fast iteration). |
AGNITRA_LLM_MODEL | Override the default Responses API deployment (e.g. gpt-5-codex). |
AGNITRA_MAX_BATCH_SIZE | Upper bound used by the runtime patcher when auto-tuning Triton kernels. |
Logging & Diagnostics
| Variable | Description |
|---|---|
AGNITRA_LOG_LEVEL | Controls SDK/CLI verbosity (INFO, DEBUG, TRACE). |
AGNITRA_PROFILER_SAMPLES | Number of warmup iterations the telemetry collector performs before capturing metrics. |
AGNITRA_BENCHMARK_REPEATS | Overrides default repeats for agnitra benchmark runs. |
Marketplace Credentials
| Provider | Environment Variables |
|---|---|
| Stripe | STRIPE_API_KEY, STRIPE_PRODUCT_ID, STRIPE_METER_ID |
| AWS Marketplace | AWS_ACCESS_KEY_ID, AWS_SECRET_ACCESS_KEY, AWS_REGION, AGNITRA_AWS_PRODUCT_CODE |
| GCP Marketplace | GOOGLE_APPLICATION_CREDENTIALS, AGNITRA_GCP_PRODUCT_ID |
skipped or deferred. Use this to rehearse workflows before enabling billing.
Feature Flags in Tests
PyTest suites monkeypatch the CLI to inject lightweight profilers and bypass heavy dependencies. Respect the following hooks when adding new commands:cli.main.profile_model— override profiler behaviour in tests.AGNITRA_TEST_MODE— signal to the SDK that fixture telemetry should be generated instead of real runs.
Further Reading
- SDK & CLI Guide for command-line equivalents of many flags.
- Telemetry Playbook to see how configuration changes affect exports.
pyproject.tomllists optional extras (openai,rl,nvml,marketplace) that unlock additional features documented here.