Using KwikQA
Bring your own key
The BYOK plans are coming soon, with the other paid plans. Until they open, every account uses Hobby, which runs on KwikQA's hosted models; this page describes BYOK as it'll work.
On the BYOK plans, the CLI sends requests straight to the model endpoint you choose, with your own key, and that endpoint bills your account for model usage. Your evidence never passes through KwikQA's servers.
| Plan | Price | Limits |
|---|---|---|
| BYOK | $19 / month or $190 / year | 200 analyses and 50 fixes a month |
| BYOK Teams | $29 / seat / month or $290 / seat / year, 3+ seats | 250 analyses and 65 fixes a month per seat, shared by the team |
| BYOK Enterprise | $49 / seat / month or $490 / seat / year, 12+ seats | 500 analyses and 150 fixes a month per seat, shared by the organization |
On BYOK Teams and BYOK Enterprise, a seat is an active committer, as on the other team plans. BYOK has the CI rules of Pro, and BYOK Teams and BYOK Enterprise have those of Teams and Enterprise. See Plans and usage and Plan rules in CI.
To count requests toward your plan, the CLI tells KwikQA each request's kind (analysis or fix), model, and provider before it sends it, and nothing else. If the endpoint doesn't answer, the request is given back.
Model endpoints
Every BYOK plan (BYOK, BYOK Teams, and BYOK Enterprise) can use any of these endpoints, with your own key. KWIKQA_ENDPOINT chooses one, from kwikqa 0.4.0:
- OpenRouter (
openrouter), the default. - Anthropic API (
anthropic), with a Claude API key. - OpenAI API (
openai). - Azure OpenAI (
azure-openai), by deployment. - Amazon Bedrock (
bedrock), with SigV4 keys or a Bedrock API key. - GitHub Models (
github-models), with a GitHub token.
Subscription logins for Claude Code, Codex, or Copilot aren't model endpoints: their terms tie them to the vendor's own apps. Use the same vendor's API instead: a Claude API key, an OpenAI API key, or GitHub Models with a GitHub token. To hold the test changes those agents make to KwikQA's gates, use kwikqa check.
1. Subscribe
$ npx kwikqa upgrade byo
$ npx kwikqa upgrade byo-team --seats 4
Or choose the plan in Billing.
2. Provide your key
Everything comes from the environment, so a CI job sets it from its own secrets. For OpenRouter, the default, export your key in the shell that runs KwikQA:
$ export OPENROUTER_API_KEY=sk-or-...
For another endpoint, set KWIKQA_ENDPOINT and that endpoint's credentials:
KWIKQA_ENDPOINT | Credentials |
|---|---|
anthropic | ANTHROPIC_API_KEY |
openai | OPENAI_API_KEY, and optionally OPENAI_ORG_ID and OPENAI_PROJECT_ID |
azure-openai | AZURE_OPENAI_ENDPOINT (https) and AZURE_OPENAI_API_KEY, and optionally AZURE_OPENAI_API_VERSION |
bedrock | AWS_REGION, and AWS_BEARER_TOKEN_BEDROCK or AWS_ACCESS_KEY_ID and AWS_SECRET_ACCESS_KEY (with AWS_SESSION_TOKEN if needed) |
github-models | GITHUB_TOKEN with models: read, and optionally KWIKQA_GITHUB_MODELS_ORG |
$ export KWIKQA_ENDPOINT=azure-openai
$ export AZURE_OPENAI_ENDPOINT=https://acme.openai.azure.com
$ export AZURE_OPENAI_API_KEY=...
$ export KWIKQA_DIAGNOSIS_MODEL=kwikqa-gpt-5 # your deployment
KwikQA reads credentials from the environment and never stores them. If one is missing, commands stop before sending anything. Your tests never see them: capture, probing, and validation run without your endpoint's model key, OPENROUTER_API_KEY, and KWIKQA_TOKEN. GITHUB_TOKEN and AWS access keys are passed through, because tests often need them, so give KwikQA a model-only credential, such as a Bedrock API key, where you can.
3. Choose models
On an endpoint other than OpenRouter, KWIKQA_DIAGNOSIS_MODEL is required: the model ID or deployment your endpoint serves, such as claude-sonnet-4-5 on the Anthropic API. KWIKQA_REPAIR_MODEL is optional; repairs use the diagnosis model unless it's set. Optionally, set KWIKQA_MODEL_CONTEXT (the context window, 128000 unless set), KWIKQA_MODEL_MAX_OUTPUT, KWIKQA_MODEL_REASONING=true for a reasoning model, and KWIKQA_MODEL_PRICING=3,15 (your USD price per million prompt and completion tokens), so the preview's cost estimate is yours.
On OpenRouter, by default you use the same models as the hosted plans. To choose your own, set routes as author/model@provider:
$ export KWIKQA_DIAGNOSIS_MODEL=anthropic/claude-haiku-5.5@anthropic
$ export KWIKQA_REPAIR_MODEL=openai/gpt-6.1-sol@openai
- On OpenRouter, the model must support strict JSON Schema output through that provider. KwikQA checks OpenRouter's endpoint details before every request and refuses routes that can't meet it.
- OpenRouter requests use only the provider you name, with fallbacks turned off.
- With your own model routes, KwikQA sends no reasoning-effort setting.
- Token and cost caps still apply to every request and are shown in each preview.
KwikQA's prompts and caps are tuned for the hosted models. Other models may diagnose less reliably.