Environment variables¶

Tulip configures itself from environment variables. The notebook harness in examples/config.py reads them with a consistent fallback chain; the SDK itself reads provider API keys directly.

This page is the answer when an example prints "skipped — missing env vars".

Minimum set for a notebook run¶

export TULIP_MODEL_PROVIDER=openai
export TULIP_MODEL_ID=gpt-4o
export OPENAI_API_KEY=sk-...

python examples/notebook_06_basic_agent.py

TULIP_MODEL_PROVIDER is the switch. Unset, it defaults to mock and the notebooks run offline against a bundled model — setting only TULIP_MODEL_ID still gives you the mock.

Model selection¶

Variable	Used by
`TULIP_MODEL_PROVIDER`	`examples/config.py:get_model` — `mock` (default), `openai`, `anthropic`
`TULIP_MODEL_ID`	Model id for the selected provider (e.g. `gpt-4o`, `claude-sonnet-4-6`)
`TULIP_MODEL_ID_B`	Secondary "model B" slot — comparison notebooks
`TULIP_MODEL_ID_C`	Tertiary "model C" slot
`TULIP_RESEARCH_MODEL`	Overrides the model for the research/eval scripts
`TULIP_NOTEBOOK_BOOT`	Set by the notebook harness during boot; not user-facing

Provider keys¶

Variable	Used by
`OPENAI_API_KEY`	`OpenAIModel`, `OpenAIEmbeddings`
`ANTHROPIC_API_KEY`	`AnthropicModel`
`COHERE_API_KEY`	`CohereEmbeddings`, `CohereReranker`

OpenAI-compatible providers¶

Every prefix in the compatible-provider table reads its own key variable — GROQ_API_KEY, TOGETHER_API_KEY, OPENROUTER_API_KEY, DEEPSEEK_API_KEY, MISTRAL_API_KEY, XAI_API_KEY, FIREWORKS_API_KEY, CEREBRAS_API_KEY, PERPLEXITY_API_KEY, NVIDIA_API_KEY — and each accepts a TULIP_<PREFIX>_BASE_URL override for self-hosted deployments and proxies. Local servers (ollama, vllm, lmstudio, llamacpp) need no key.

Storage backends¶

Variable	Used by
`REDIS_URL`	Redis checkpointer — notebooks 08, 68
`S3_BUCKET`	S3 checkpointer — notebook 52
`S3_ENDPOINT_URL`	S3-compatible endpoint (MinIO, R2) — notebook 52

Serving¶

Variable	Used by
`TULIP_SERVER_API_KEY`	`AgentServer` bearer token — notebook 68
`TULIP_A2A_API_KEY`	`A2AServer.__init__` bearer token

LiteLLM gateway¶

Notebooks 71 and 72. See the gateway how-to.

Variable	Used by
`LITELLM_GATEWAY_URL`	Gateway base URL
`LITELLM_GATEWAY_KEY`	Virtual key the agent presents
`LITELLM_GATEWAY_MODEL`	Model id as the gateway names it
`LITELLM_MASTER_KEY`	Admin key — `/key/generate` and `/spend/*` (needs Postgres)

Cloud and security domain¶

Variable	Used by
`TULIP_AWS_PROFILE`	AWS cloud-posture tools — notebook 73
`AWS_ACCESS_KEY_ID`	Standard AWS credential chain
`AWS_SECRET_ACCESS_KEY`	Standard AWS credential chain
`SCANNER_API_KEY`	Vulnerability-scanner adapter — notebook 70
`DATAMAP_URL`	Data-map adapter — notebook 70
`DATAMAP_API_KEY`	Data-map adapter — notebook 70
`DATAMAP_TOKEN`	Data-map adapter — notebook 70

Model fingerprinting¶

Notebook 80. See cloud posture.

Variable	Used by
`FINGERPRINT_MODEL`	Model id under test
`FINGERPRINT_ASSET`	Asset label recorded on the finding
`FINGERPRINT_SAMPLES`	Number of timing samples to collect

Other¶

Variable	Used by
`OPENCLAW_DIR`	OpenClaw workspace root — the red-team example