Bedrock — Amazon Bedrock Runtime
Quick start, the two Runtime operations and the modelId forms accepted, and what a canned response means for token counts, latency and streaming.
Converse and InvokeModel answer from a canned response. No model runs, so
this exercises your wiring — not your prompts.
Status: ⚠️ Partial
Quick start
export AWS_ENDPOINT_URL=http://localhost:4566
aws bedrock-runtime converse \
--model-id anthropic.claude-3-5-sonnet-20241022-v2:0 \
--messages '[{"role":"user","content":[{"text":"hello"}]}]'Any credentials work; with none configured, run eval "$(overcast env)" first
— see Using AWS SDKs and CLI.
The reply is a complete ConverseResponse — every member the AWS model marks
required, with fixed assistant text in output.message.content[0].text.
What works
| Area | Behaviour |
|---|---|
Converse | POST /model/{modelId}/converse; full response shape, so an SDK deserialises it without complaint |
InvokeModel | POST /model/{modelId}/invoke; answers {"overcastEmulator": …} |
modelId | A model ID, an inference-profile ID, or an ARN — percent-encoded or not |
Differences from AWS
| Area | On AWS | Overcast |
|---|---|---|
| Inference | The named model runs | Nothing runs; the text is fixed regardless of the prompt |
| Usage metrics | Real token counts and metrics.latencyMs | All zero |
InvokeModel body | The model’s own schema | One overcastEmulator field |
| Streaming | InvokeModelWithResponseStream, ConverseStream | Not implemented — 501 Not Implemented |
CountTokens | Counts tokens for a model | Not implemented — 501 Not Implemented |
Gotchas
Tip
InvokeModel’s body is opaque on AWS too — its schema belongs to the model,
not to Bedrock. Use Converse where you need a response shape AWS itself
defines.
Operations
All 2 listed operations are implemented. Per-operation status, notes and AWS API links: Bedrock operations.