Task-shaped: you already know what you want, and this tells you how. If you are still working out what you want, read an explanation page instead; if you know the name of the setting and want its default, that is reference.
Begin with the quickstart for a running stack. Individual guides record their own examples and validation; a documentation refresh does not represent a new live-provider run.
- Configure Pointsman - environment, YAML and runtime records; worked examples, activation, verification and rollback.
- Prepare a deployment - network boundaries, persistence, resource sizing and rollout validation.
- Add a provider and a model - any OpenAI-compatible vendor as a database row: create disabled, probe, enable, prove a request landed there.
- Add a second key to a provider, and rotate one - a sibling credential, the ledger proof of which key paid, a rotation with a stream in flight, a per-key ceiling, and every refusal the door makes.
- Build a team hierarchy - an authored team scope, a correctly caged team admin, a key homed at the team, and a budget whose binding the server names.
- Issue a key and give it a budget - the admin login dance, an organisation, a monthly cap, and a key that draws on it.
- Send some of your traffic to a different model - canary deploys, A/B tests and gradual migrations, which are the same mechanism with different numbers.
- Turn on outcome learning - the bandit strategy: decision ids, reporting how answers turned out, and reading the belief surface while it learns.
- Route by task with an external classifier - the opt-in Jev strategy: what it exports and to whom, the operator-paid bounds, the ranking it applies inside the shortlist, the fallback on every abstention, and the offline evaluation to run before trusting it.
- Route only to models with quality evidence - the experimental quality-guarded strategy: register validators, freeze an evaluation, adjudicate one sealed batch, preview and promote a policy, read what served each request, roll back without reviving expired evidence.
- Write and enforce a guardrail - dry-run against sample text, watch in monitor mode, then enforce; including the misconfigurations made writing it.
- Tune the semantic cache - the similarity threshold, the evidence to tune it from, and what the cache refuses to do.
- Work out what happened to a request - from a request id, or from the queryable record when you do not have one.