Your gateway (InvokeModel)

Many organisations reach Claude through their own gateway: plain HTTPS, an API-key header, and the Bedrock InvokeModel operation. bedrock() signs with AWS credentials against the Converse operation, and providerFromEnv() has no base URL for this wire — so neither reaches it. invokeModelGateway() does, with no SDK.

Use

const provider = withRetry(  invokeModelGateway({    baseUrl: 'https://llm-gateway.example.com/bedrock',    apiKeyHeader: 'api-key',    // One key per model: this gateway scopes keys, and the wrong one is a 403.    apiKey: { [HAIKU]: 'haiku-key', [SONNET]: 'sonnet-key' },    model: HAIKU,    fetch: scriptedGateway(log), // ← drop this line to talk to the real gateway  }),); // a 429, a 5xx or a network failure is retried by withRetry, never inside the adapter

Then Agent.create({ provider, model: 'invoke-model-gateway' }). Prefer new? new InvokeModelGatewayProvider(options) is the same provider as a class. The shorthand 'invoke-model-gateway' sends the model you configured; any other model name is sent as given, so one provider can serve several models.

The options object is typed InvokeModelGatewayOptions:

A key per model

Gateways often scope a key to some models, and asking with the wrong one gets a 403. Give each model id its own key with a map (as above) or a function. A model the map does not name is refused before any request is sent, and the error lists the ids the map does name. The key itself is never in an error message.

What goes on the wire

The body is Anthropic's Messages API, with anthropic_version: "bedrock-2023-05-31" and no model field (the model is in the path). The system prompt rides the top-level system field; tool calls go out as tool_use blocks and results come back as tool_result blocks on a user turn. A forced tool (tool_choice) is supported.

stream() reads the SSE the gateway sends after decoding AWS's binary framing. Bare data: lines and event: + data: pairs both work. Text arrives as chunks; tool arguments are reassembled from their input_json_delta fragments, and the final chunk carries the full response (tool calls, usage, stop reason).

Errors

Every failure is an InvokeModelGatewayError with a reason (typed InvokeModelGatewayErrorReason) and the modelId:

"Prompt is too long" becomes a ContextWindowExceededError instead, and a cancelled request is passed through as the AbortError it is.

Retries: compose withRetry

The provider makes one attempt per call and has no retry loop of its own. Every InvokeModelGatewayError carries retryable: true for a 429, a 5xx and a network failure, false for everything else. withRetry's default policy honours the false, so it retries exactly those three:

• A 403 or 404 is an answer, not a passing fault, and fails straight away.
• A refusal raised before any request (no-key, no-model) is never repeated, and nothing waits out a backoff for it.
• A 2xx the provider could not read (unreadable-response) is never re-sent: the model may already have run, and billed.

const log: string[] = [];const provider = withRetry(  invokeModelGateway({    baseUrl: 'https://llm-gateway.example.com/bedrock',    apiKeyHeader: 'api-key',    apiKey: 'sonnet-key',    model: SONNET,    fetch: scriptedGateway(log, true),  }),  { initialDelayMs: 10 },);await provider.complete({  model: 'invoke-model-gateway',  messages: [{ role: 'user', content: 'hi' }],});

Next steps

• AWS Bedrock — bedrock(), for AWS credentials and the Converse API
• Resilience — withRetry / withFallback
• Custom provider — createProvider({ kind: 'invoke-model-gateway', ... })