Control every model. Keep every team moving.

It chooses the best model for each task, not one global default. Every candidate is scored on five weighted axes — capability 35%, cost 25%, latency 20%, reliability 15%, org preference 5% — gated by your policy. A bulk refactor routes to a fast, cheap model; an architectural review routes to a frontier reasoning model. You tune the weights and pin models per task type in policy, and the full decision (per-model scores, cost, reason) is written to the audit trail. Nothing is a black box.

Yes — multi-model agent workflows are first-class. A run can use a reasoning model to orchestrate and plan (e.g. claude-fable-5), a fast model to implement each step (e.g. claude-sonnet-4-6), and a separate strong model to review the diff (e.g. gpt-5.5). Each role is configurable in policy — swap any model, pin a provider, or set approval gates between steps. Every step is logged with its own cost and trace.

No. Policate speaks the Anthropic API, so any tool that already does — our improved agent CLI, Pi, Claude Code, the Anthropic SDK — just points at the gateway. Set ANTHROPIC_BASE_URL and ANTHROPIC_API_KEY (or install the policate CLI) and every request is governed with under 5ms of overhead. With our CLI you additionally control approved Skills, slash Commands, MCP servers, and Hooks org-wide — pushed to every developer automatically, with per-developer install status visible to admins. Keep your own tool if you prefer; you still get full routing, policy, cost, and audit.

Bring your own keys — Bedrock IAM roles, Anthropic, OpenAI. They are encrypted at rest, stored as hashes, and never returned in an API response or written to a log. Policate tokens themselves are stored only as their sha256, and the raw token is shown exactly once on creation.

Data residency is a first-class policy control. Route to EU-only cross-region inference profiles (for example eu.anthropic.claude-sonnet-4-6) and the gateway enforces it before the request leaves the machine. Residency requirements are versioned in policy alongside model, tool, and budget rules.

Every request is metered and attributed by org, team, developer, and model. Budgets are enforced atomically by strongly consistent counters, so a request is never double-counted and hard blocks land the moment a cap is hit. Cache hits cost nothing and are reported as such — savings attribution is in the dashboard from week one.