Amazon Bedrock Mantle Provider Setup for OpenClaw

OpenClaw ships with an Amazon Bedrock Mantle provider built in, and it talks to Mantle's OpenAI-compatible endpoint. Through a standard /v1/chat/completions surface running on Bedrock infrastructure, Mantle serves open-source and third-party models such as GPT-OSS, Qwen, Kimi, GLM, and others of that kind. Anthropic Claude models are also available on Mantle, reachable via an Anthropic Messages route.

Getting started

Pick whichever auth method suits you and work through the setup steps.

Explicit bearer token

Best for: environments where a Mantle bearer token is already on hand.

Set the bearer token on the gateway host

export AWS_BEARER_TOKEN_BEDROCK="..."

You can also set a region if you want something other than the default of us-east-1:

export AWS_REGION="us-west-2"

Verify models are discovered

openclaw models list

Models that get discovered show up under the amazon-bedrock-mantle provider. Nothing else needs configuring unless you want to change the defaults.

IAM credentials

Best for: AWS SDK-compatible credentials (shared config, SSO, web identity, instance or task roles).

Configure AWS credentials on the gateway host

Any auth source that works with the AWS SDK is acceptable:

export AWS_PROFILE="default"
export AWS_REGION="us-west-2"

Verify models are discovered

openclaw models list

From the credential chain, OpenClaw mints a Mantle bearer token on its own.

Tip

If AWS_BEARER_TOKEN_BEDROCK has not been set, OpenClaw creates the bearer token for you, drawing on the AWS default credential chain: shared credentials and config profiles, SSO, web identity, and instance or task roles all count.

Automatic model discovery

With AWS_BEARER_TOKEN_BEDROCK set, OpenClaw takes it as-is. If it is not set, OpenClaw tries to mint a Mantle bearer token from the AWS default credential chain. After that it queries the region's /v1/models endpoint to find which Mantle models are available.

When a refresh fails, the report says catalog access is unavailable or rejected. The catalog retains the compatible last-good models alongside that failure status; discovery data that has expired never turns into a successful refresh. A successful response with no models clears discovered membership. Once endpoint access is restored, refresh once more.

The plugin's public discovery helpers keep their v2026.9.2 advisory defaults: on failure, stale rows from the same credential may still be returned, and implicit resolution yields null for empty results. Programmatic callers may pass discoveryMode: "strict" to propagate failures and to get successful empty provider results. The bundled catalog hooks always pick that strict mode.

To leave the Mantle plugin on while turning off automatic discovery and IAM bearer-token generation, flip off the plugin-owned discovery toggle:

openclaw config set plugins.entries.amazon-bedrock-mantle.config.discovery.enabled false

Note

This bearer token is the very same AWS_BEARER_TOKEN_BEDROCK that the standard Amazon Bedrock provider uses.

Supported regions

us-east-1, us-east-2, us-west-2, ap-northeast-1, ap-south-1, ap-southeast-3, eu-central-1, eu-west-1, eu-west-2, eu-south-1, eu-north-1, sa-east-1.

Manual configuration

Should you want explicit config rather than auto-discovery:

{
  models: {
    providers: {
      "amazon-bedrock-mantle": {
        baseUrl: "https://bedrock-mantle.us-east-1.api.aws/v1",
        api: "openai-completions",
        auth: "api-key",
        apiKey: "env:AWS_BEARER_TOKEN_BEDROCK",
        models: [
          {
            id: "gpt-oss-120b",
            name: "GPT-OSS 120B",
            reasoning: true,
            input: ["text"],
            cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 },
            contextWindow: 32000,
            maxTokens: 4096,
          },
        ],
      },
    },
  },
}

Membership is governed by an explicit non-empty models list, which replaces discovered rows, the Claude rows below included. Where rows match, an explicit input takes precedence; if the source row leaves out input, discovery is allowed to supply that capability metadata. Leave models out to keep the automatic Mantle catalog, or list every Claude model entry you intend to use.

Advanced configuration

Reasoning support

Reasoning support gets inferred from model IDs that contain patterns such as thinking, reasoner, reasoning, deepseek.r, gpt-oss-120b, or gpt-oss-safeguard-120b. During discovery, OpenClaw assigns reasoning: true on its own for models that match.

Endpoint unavailability

Requests that fail against the endpoint report catalog access as unavailable or rejected. Compatible last-good models stay listed with that failure status. A successful response with no models clears discovered membership. When no bearer token can be resolved, discovery never runs and the implicit provider is skipped. Any other configured providers keep working as usual.

Claude via the Anthropic Messages route

When the model list is managed by automatic discovery, OpenClaw adds five Claude models once an /v1/models lookup succeeds and returns one or more models: amazon-bedrock-mantle/anthropic.claude-opus-5 (Claude Opus 5), amazon-bedrock-mantle/anthropic.claude-sonnet-5 (Claude Sonnet 5), amazon-bedrock-mantle/anthropic.claude-opus-4-7 (Claude Opus 4.7), amazon-bedrock-mantle/anthropic.claude-mythos-5 (Claude Mythos 5), and amazon-bedrock-mantle/anthropic.claude-mythos-preview (Claude Mythos Preview). These run on the anthropic-messages API surface and stream via the same bearer-authenticated Anthropic-compatible endpoint (<mantle-base>/anthropic), which means the AWS bearer token is not handled as an Anthropic API key.

The context window for Claude Opus 5 is 1,000,000 tokens, with a 128,000-token output limit, image input support, and $5/$25 pricing for input and output. Adaptive thinking is on by default (high); thinking can be turned off with /think off, while /think xhigh|max selects the model's native effort levels. Any sampling parameters chosen by the caller are dropped by OpenClaw.

Claude Sonnet 5 relies on adaptive thinking at all times and starts at high effort. Because the Mantle route offers no way to disable thinking, /think off and /think minimal are mapped to low. Custom temperature is likewise dropped by OpenClaw for Sonnet 5 requests.

Access to Claude Mythos 5 is limited. Its published context window is 1,000,000 tokens with a 128,000-token output limit; it always uses adaptive thinking, maps /think off and /think minimal to low, and drops caller-selected sampling parameters.

Claude Mythos Preview always asks for reasoning, and when no /think level is set it defaults to high effort (xhigh/max map down to high, while minimal maps up to low). On Mantle, Opus 4.7 streams with no model-provided reasoning, and since Opus 4.7 does not accept sampling overrides on this route, OpenClaw drops its temperature parameter; a temperature override is accepted normally by Mythos Preview.

Membership is governed by a non-empty explicit models.providers["amazon-bedrock-mantle"].models list, which takes the place of the entire discovered catalog. Discovered input capability can be inherited by matching rows only if the source row left it out. Leave the list out to keep these built-in Claude rows.

Relationship to Amazon Bedrock provider

The standard Amazon Bedrock provider and Bedrock Mantle are distinct providers. For its OSS catalog, Mantle exposes an OpenAI-compatible /v1 surface, whereas the standard Bedrock provider relies on the native Bedrock Converse API.

Where it exists, the same AWS_BEARER_TOKEN_BEDROCK credential is used by both providers.

Related

• Amazon Bedrock, Native Bedrock provider for Anthropic Claude, Titan, and other models.

• Model selection, Choosing providers, model refs, and failover behavior.

• OAuth and auth, Auth details and credential reuse rules.

• Troubleshooting, Common issues and how to resolve them.