Connect providers
Bring your own provider keys so runs use your accounts and rates.
1. Provider API keys
RedCrown stores your keys encrypted at rest and resolves them per run. Keys are scoped to your workspace and never shared across workspaces.
Go to Models & keys (/advanced/models), one click from the primary nav. Choose a provider from the list, paste your API key, and click Save. The key is stored encrypted and is never shown again after saving.
Set the provider API key as an environment variable in your shell before running redcrown eval. Keys never leave your machine during a local eval.
export OPENAI_API_KEY=sk-...
redcrown eval task.jsonSee the CLI reference for the full list of accepted environment variable names per provider.
Call list_models with no arguments to see which providers are currently connected for your workspace:
{}The response lists every available model grouped by provider. A provider that appears here has a connected key.
2. Free provider tiers, no credit card
You can connect a real model and run a real eval without a credit card. Six providers offer a no-card free tier that can finish a real eval today (Hugging Face also has a no-card free tier, but its $0.10 a month cannot finish one, so it is not in the table below). Sign up and paste the key at Models & keys (/advanced/models) or export it as an environment variable for the CLI.
Groq and OpenRouter models already appear in the model picker once you connect a key: RedCrown carries curated free presets for both. Google AI Studio (Gemini), NVIDIA NIM, Mistral, and Cohere are connectable today too, but RedCrown does not yet carry curated presets for those four, so until they are added to the catalogue, you name the model yourself, either directly in an experiment spec or through a custom endpoint.
| Provider | Sign up | Ceiling |
|---|---|---|
| Google AI Studio (Gemini) | aistudio.google.com/apikey | Free tier on Gemini Flash and Flash-Lite models, per-project rate limits shown in AI Studio |
| NVIDIA NIM | build.nvidia.com | About 1,000 inference credits on joining the developer program, about 40 requests per minute |
| Groq | console.groq.com/keys | Free plan rate limits per model, shown on the console rate-limits page once signed in |
| OpenRouter | openrouter.ai/keys | Models ending :free run at a $0 balance, 20 requests per minute and 50 requests per day until $10 of credit has ever been purchased |
| Mistral | console.mistral.ai | Free Experiment tier, rate limited, phone verification required at signup |
| Cohere | dashboard.cohere.com/api-keys | Trial key, 20 requests per minute on chat and 1,000 API calls per month, evaluation use only |
Free is not unlimited. Every row above is rate limited, and the Gemini and Mistral free tiers may use what you send them to improve their models, so data under a privacy or compliance obligation belongs on a paid tier, not a free one.
Cloudflare Workers AI is not a registry provider, but its OpenAI-compatible API gives 10,000 neurons a day free on any account and can be connected as a custom endpoint. Use your Cloudflare account id in the base URL and a Workers AI model name:
Base URL: https://api.cloudflare.com/client/v4/accounts/{account_id}/ai/v1
Auth header: Authorization: Bearer <cf-token>
Body style: openai
Model: @cf/meta/llama-3.1-8b-instructTwo options that show up in "best free LLM API" roundups are deliberately absent from this list. GitHub Models was fully retired on 2026-07-30, its API and BYOK are both off. Cerebras now requires a verified payment method before it enables API access, so it is not a no-card option.
3. AWS Bedrock
Bedrock requires an AWS access key, secret, and region. RedCrown stores the three fields as an encrypted credential blob.
Go to Models & keys (/advanced/models). Select AWS Bedrock, then enter your Access key ID, Secret access key, and Region (for example, us-east-1). Click Save.
The standard AWS environment and credential chain applies. Set the environment variables or configure a profile with the AWS CLI before running redcrown eval:
export AWS_ACCESS_KEY_ID=AKIA...
export AWS_SECRET_ACCESS_KEY=...
export AWS_DEFAULT_REGION=us-east-1
redcrown eval task.jsonNot available via MCP. Connect Bedrock in the app.
4. A custom endpoint
A custom HTTP endpoint lets you register any authenticated API that follows an OpenAI-compatible or plain request/response shape as a candidate model.
Go to Models & keys (/advanced/models). Select Custom endpoint and fill in the Base URL, the Model it serves (for example gemini-2.5-flash), the Auth header (for example, Authorization: Bearer ...), and the Body style (openai or raw). Click Save.
The Model field is required. An endpoint saved without one is stored and shown as connected, but no evaluation can call it, because nothing knows what model to ask for. Only one custom endpoint is stored per workspace, so saving replaces the one already there. Your endpoint is offered when you pick models yourself; it is not added automatically to a scaffolded run.
Set REDCROWN_CUSTOM_HTTP to a JSON object with the connection details to run a custom_http candidate locally:
export REDCROWN_CUSTOM_HTTP='{"base_url":"https://my-model.example.com","auth_header":"Bearer my-token","body_style":"openai","model":"my-model-v1"}'
redcrown eval task.jsonNot available via MCP. Use the UI or the CLI environment variable above.
5. What happens when a key is missing
A run whose models RedCrown cannot reach is refused before it starts, rather than run to completion with every item failing. Starting a run returns 400 with a message naming what to connect:
None of this experiment's models can run: no key is connected for
anthropic, openai. Connect a key at /connections, or pick models that
match your data.Three cases are refused this way:
- No key for any candidate. Connect one of the named providers. A run where some candidates are reachable is still allowed: those get ranked, and the unreachable ones show as errored rows in the report.
- A speech model over a text dataset. Transcription models take audio, so they cannot score pasted text. Upload audio, or pick text models for that step.
- Judge scoring with no reachable judge. The
judgeandrubricquality metrics grade with an LLM. With no key for the judge's provider nothing can be scored at all, so connect that key or choose a deterministic metric such asexact_matchorsimilarity.
If you have no key connected at all, RedCrown may instead run a free sample on models it pays for. That report and its shareable proof say so at the top, and name the configs you chose that were not tested. A sample run never claims to have measured your own models.
Last verified .