@tailored-ai/provider-bedrock

AWS Bedrock as a model provider. One provider id, bedrock, covers every model your AWS account can invoke: Anthropic Claude, Amazon Nova, Meta Llama, Mistral, and the rest of the Bedrock catalog. The plugin speaks Bedrock's Converse API, which normalizes message and tool-calling shapes across model families, so agents and their tools work the same regardless of which model you pick.

This is also the reference implementation for a provider plugin: a single register(ctx) package that adds a new entry to the provider registry. If you want to write your own provider, start from this package's source.

Install

bash

tai plugin install @tailored-ai/provider-bedrock

Configure

Three pieces: load the plugin, describe the provider, select it.

yaml

plugins:
  - "@tailored-ai/provider-bedrock"

providers:
  bedrock:
    defaultModel: "us.anthropic.claude-haiku-4-5-20251001-v1:0"
    region: us-west-2        # optional: falls back to AWS_REGION / profile config
    profile: my-profile      # optional: falls back to the default credential chain

agent:
  defaultProvider: bedrock

Or keep your local model as the default and give a single agent a Bedrock model:

yaml

agents:
  researcher:
    provider: bedrock
    model: "us.amazon.nova-pro-v1:0"

Credentials

No keys in config.yaml. The plugin uses the standard AWS credential chain: environment variables, ~/.aws profiles, SSO sessions, and instance roles. profile pins a named profile when you have several.

The IAM principal needs bedrock:InvokeModel on the models you use. Check your access from the same machine TAI runs on:

bash

aws sts get-caller-identity
aws bedrock list-foundation-models --region us-west-2

Model ids

Most current Bedrock models reject on-demand invocation under their bare model id and require a cross-region inference profile id instead. If you see Invocation of model ID ... with on-demand throughput isn't supported, add your region-group prefix:

text

anthropic.claude-haiku-4-5-20251001-v1:0      ✗
us.anthropic.claude-haiku-4-5-20251001-v1:0   ✓

Accounts homed in Europe or Asia-Pacific use eu. / apac.. List the profile ids your account can invoke:

bash

aws bedrock list-inference-profiles --region us-west-2 \
  --query 'inferenceProfileSummaries[].inferenceProfileId'

Model-specific request fields

Anything in ChatParams.extra passes through as Converse's additionalModelRequestFields. That is the escape hatch for per-family knobs (Anthropic's top_k, Nova's inferenceConfig extensions) without the plugin needing to know about them.

How it maps

TAI	Converse
Leading `system` messages	Top-level `system` blocks
Assistant `toolCalls`	`toolUse` content blocks
`tool` role results	User-turn `toolResult` blocks
`finishReason: tool_calls / length / stop`	`stopReason: tool_use / max_tokens / end_turn`

Adjacent same-role messages merge into one turn (Converse requires alternating user/assistant turns), and empty messages are dropped (Converse rejects empty text blocks).

Streaming

The provider implements chatStream via ConverseStream, so chat responses render incrementally in the web UI. Text arrives as deltas; tool calls are accumulated provider-side and surface complete. The IAM principal needs bedrock:InvokeModelWithResponseStream alongside bedrock:InvokeModel — without it, streaming calls fail while plain chat keeps working.