How it works
Claude Code's built-in advisor lets the working model consult a stronger model at decision points, but it runs as a server-side tool on Anthropic's API and isn't available when Claude Code is configured for Amazon Bedrock. claude-advisor re-creates the same workflow locally, outside the API.
Opus orchestratorspawns a subagent
Sonnet / Haiku subagentruns the task
advisor toolcalled at a decision point
plugin hookfinds the agent's transcript on disk
advisor-enginelocal Go binary, filters transcript
Bedrock (Opus)Converse API
Guidancereturned as the tool result
The plugin (mod) is named bedrock-advisor and registers the advisor
tool. It can also consult automatically, without the subagent asking: if the same tool error
repeats twice in a row, the plugin calls the advisor and attaches the guidance to the tool
result.
Requirements
- Claude Code v2.1.293 or later.
- Claude Code configured for Bedrock (
CLAUDE_CODE_USE_BEDROCK=1). - For the direct path: AWS credentials the AWS SDK default chain can find (profile, SSO, or environment variables),
AWS_REGION, and Bedrock access to the advisor model. Without them, the defaultautobackend sends the advisor request through Claude Code's own model call instead.
Configuration
Set with /config, or from your shell with echo '{"advisor_model":"sonnet"}' | claude plugin configure bedrock-advisor@model-harness --values-stdin.
| Option | Default | Description |
|---|---|---|
advisor_model |
opus |
Alias (opus, sonnet, haiku) resolved through ANTHROPIC_DEFAULT_OPUS_MODEL and the equivalent Sonnet/Haiku variables, or a full Bedrock model / inference-profile ID. |
backend |
auto |
auto: the engine calls Bedrock directly and falls back to Claude Code's own model call on credential or access-denied errors. bedrock: engine only. session: Claude Code's model call only. |
auto_consult |
true |
Consult the advisor automatically when the same tool error happens twice in a row. |
max_tokens |
2048 |
Cap on the advisor's reply length. |
max_context_chars |
400000 |
Transcript budget sent to the advisor. |
engine_path |
— | Path to the advisor-engine binary. Set by the installer. |
Compared with the built-in advisor
| Built-in advisor | claude-advisor | |
|---|---|---|
| Execution | Runs inside one API request, server-side | Runs as a normal tool call — one extra round trip |
| Platform | Anthropic API only | Works on Bedrock |
| Context source | Shares the request's context server-side | Reads the agent's transcript from disk |
| Repeated errors | No setting to force a call | Can force a consultation when the same tool error repeats |
| Advisor prompt | Anthropic's, not published | Open source, in the engine |
| Billing | Advisor model rates on the Anthropic API | Advisor tokens billed by Bedrock at the advisor model's rates |