Agree the setup before you switch.
This is a managed setup and migration service, quoted for your application. There is no public Flying Penguin API endpoint, self-service key dashboard or token subscription available from this website today.
Before production use, we agree and test the regional route, compatibility, access, usage controls, retention, support and recovery arrangements. You get a written scope and the results for the setup you are buying.
Keep the model call in London.
The proposed route uses Amazon Bedrock In-Region in Europe (London), eu-west-2. AWS currently lists Claude Sonnet 4.6 and Claude Opus 4.6 for that route.
Model access and capacity need checking in the actual account. An EU or global inference profile is a different arrangement; it does not meet a London-only requirement.
Bring the integration you already have.
We review your Anthropic Messages API calls and scope the gateway connection around them. The handover names the endpoint, key, model identifiers and supported features.
Ordinary messages, streaming, tool calls, errors and usage reporting need testing with your app. Some provider-specific features need changes; I will identify those before quoting the migration.
Stop rather than change the location.
For an agreed London-only service, an unavailable model or region must produce an error instead of silently switching to another region or provider. The setup needs an explicit model allow-list, restricted regional permissions and failure tests.
We keep the results with the operating notes and review them when models change.
Give each app its own access and limits.
The scope can include separate keys, expiry and revocation, IP restrictions and permissions for each application or client. We agree rate limits, spend warnings and enforced request limits separately.
Hard stops need checks for simultaneous requests, streaming and retries, as well as a simple request that goes over the limit.
See what the model use costs.
The proposal separates the initial integration, continuing management, infrastructure and model usage. Input, output and cache usage can have different prices.
We agree per-app or per-client reporting, the billing currency, any allowance and the calculation behind it. A supplier account billed in dollars is identified as such; a pounds-based management quote does not remove exchange-rate exposure.
Reduce the information sent.
Where needed, the project can include text redaction or reversible replacement of selected identifiers before the model call. We agree what to detect, how replacements are restored and how temporary mappings are removed.
We test missed identifiers, altered answers and streamed output. This is a scoped feature, not an existing universal filter; images and attachments need their own handling.
Reversible replacement still needs personal-data safeguards.
Put the data handling in writing.
We document the gateway, model, storage, operational logs, backups and support access, alongside processor responsibilities and supplier terms. Prompt bodies should be excluded from ordinary operational logs unless an agreed reason and retention policy require them.
AWS retention rules depend on the chosen model and settings, so a blanket zero-retention promise would be misleading. London processing alone does not establish UK GDPR compliance or remove every possible international transfer.