OpenCode with AWS Bedrock

This guide puts CT-Toolkit between OpenCode and your Bedrock models. Every tool call the model asks OpenCode to run is reviewed by the tool-call guard and recorded in the provenance chain, and the records appear in the web dashboard.

OpenCode  ──►  CT-Toolkit proxy :8788  ──►  LiteLLM :4000  ──►  AWS Bedrock
                 (review, record)          (OpenAI-compatible)
                        │
                        └──► provenance vault ◄── dashboard :8001

OpenCode's built-in Bedrock provider talks the Bedrock Converse protocol directly. The CT-Toolkit proxy reads OpenAI Chat Completions and Anthropic Messages, so Bedrock is reached through LiteLLM, which exposes it as an OpenAI-compatible API. Nothing in OpenCode's code changes; only its configuration does.

1. Sign in to AWS

Your AWS session must be valid for LiteLLM to call Bedrock:

aws login --profile <your-profile>

2. Start the chain

The example files are in examples/opencode-bedrock/:

cd examples/opencode-bedrock
CT_AWS_PROFILE=<your-profile> CT_MODE=observe ./start.sh

This starts three processes and prints their addresses:

Process	Address	Role
LiteLLM	http://127.0.0.1:4000/v1	Bedrock as an OpenAI-compatible API (litellm_config.yaml)
CT-Toolkit proxy	http://127.0.0.1:8788/v1	Reviews and records tool calls (--tool-guard-mode)
Dashboard	http://127.0.0.1:8001/dashboard	Shows the records from the same vault

Logs go to examples/opencode-bedrock/logs/.

3. Point OpenCode at the proxy

Add the provider from examples/opencode-bedrock/opencode.json to your OpenCode configuration (~/.config/opencode/opencode.json) and select its model:

"provider": {
  "ct-bedrock": {
    "npm": "@ai-sdk/openai-compatible",
    "options": { "baseURL": "http://127.0.0.1:8788/v1", "apiKey": "<the value of examples/opencode-bedrock/.litellm_master_key>" },
    "models": { "global.openai.gpt-6-luna": { "name": "GPT-6 Luna (Bedrock, guarded)", "tool_call": true } }
  }
}

and "model": "ct-bedrock/global.openai.gpt-6-luna". Use the same Bedrock model id in litellm_config.yaml.

4. See the records

Work in OpenCode as usual, then open the dashboard. The Provenance Chain Ledger lists every request that passed through the proxy: model requests as model turns, other API calls as API calls, and each tool call the model asked for, with its verdict (passed, flagged, blocked) and findings. Filter by kind, search, and sort by time, kind or score. The Tool calls section shows the tool-call decisions alone, with filters by verdict and tool.

The Tool-Call Guard Settings panel changes the mode without a restart: the proxy reads the settings file on every request. Enforce mode removes high-severity calls before OpenCode runs them.

What is recorded, and what is not

• Recorded: every request that passes through the proxy. A model request gets one entry with its endpoint, status, finish reason and number of tool calls, including a plain answer or a refusal. Any other request, such as a model listing, gets an entry with its method, endpoint and status.
• Recorded: each tool-call decision, with the tool name, verdict, mode, findings, endpoint and model. The call's arguments are not stored; the chain keeps only their hash. The model's text is hashed the same way.
• Not recorded yet: identity drift. Drift is measured by the library's TheseusWrapper, which the proxy does not wrap.

Verification status

• The proxy, the dashboard and the shared vault were tested end to end with a stand-in OpenAI-compatible upstream that streams tool calls, including a risky one. The risky call was flagged and recorded, and it appeared in the dashboard API.
• The Bedrock leg was run with a valid AWS session. A plain request and a request that asks for a tool call both returned Bedrock responses through the chain. The tool call was recorded in the dashboard with the verdict passed.

Enforce mode with OpenCode

In enforce mode the blocked call is removed from the stream: the text around it is still sent, and OpenCode never receives the call, so it cannot run it. Start with observe mode until you have looked at the records for a day or two.