Connect Amazon Bedrock

For operators who self-host. How-to.

Drafts come from Amazon Bedrock (Claude Haiku 4.5) through one gated tool. Until you turn this on, drafts come from an offline template. A person always edits a draft and posts it themselves.

What you need first

• An AWS account with access to the model. Submit the one-time Anthropic use-case form in the Bedrock console.
• Live mode. See Replay mode and live mode.
• Credentials the server can find. On the AWS host that is the instance role.

Steps

1. Check that the model answers from the same account:

   aws bedrock-runtime converse --model-id us.anthropic.claude-haiku-4-5-20251001-v1:0 \
     --messages '[{"role":"user","content":[{"text":"Say ok."}]}]'

2. Set the region and model id if yours differ from the defaults:

   CSO_BEDROCK_REGION=us-east-1
   CSO_BEDROCK_MODEL_ID=us.anthropic.claude-haiku-4-5-20251001-v1:0

3. Turn on the gate. A real Bedrock call needs both CSO_MODE=live and CSO_LIVE_BEDROCK=1:

   CSO_MODE=live
   CSO_LIVE_BEDROCK=1

   With either one missing, drafts use the offline template and no AWS client is created.

4. Set the spend guard. The default is 10 US dollars per UTC day:

   CSO_BEDROCK_DAILY_USD_CAP=10

   Past the cap, drafts fall back to the template. CSO_DAILY_DRAFT_CAP (default 100) limits the number of drafts per day. See Replay mode and live mode for how the caps count.

5. Restart the API and the worker.

On the AWS host, do not edit /opt/cso/.env by hand. infra/scripts/bootstrap.sh rewrites that file from SSM Parameter Store each time it runs, so a hand edit is lost. Set the gate with an optional parameter instead. It is a switch, not a secret, so a plain String is right:

aws ssm put-parameter --name /cso/prod/live_bedrock --type String --value 1 --overwrite

Use your own stage in place of prod. Then run the bootstrap script again through SSM Session Manager. It is safe to repeat, it writes CSO_LIVE_BEDROCK=1 into /opt/cso/.env and it restarts the stack:

aws ssm start-session --target "$ID"
sudo bash /opt/cso/infra/scripts/bootstrap.sh

The host runs in live mode already (CSO_MODE=live), so the second half of the gate is set. Only the exact value 1 turns drafts on. Delete the parameter (or set it to 0) and run the script again to go back to offline templates. The instance role already allows bedrock:InvokeModel for the two ARNs under IAM below.

IAM

The server needs bedrock:InvokeModel on two resources only: the inference profile and the foundation model for Claude Haiku 4.5. The AWS template gives the instance role exactly that:

{"Effect":"Allow","Action":["bedrock:InvokeModel"],
 "Resource":["arn:aws:bedrock:us-east-1:<account id>:inference-profile/us.anthropic.claude-haiku-4-5-20251001-v1:0",
             "arn:aws:bedrock:*::foundation-model/anthropic.claude-haiku-4-5-20251001-v1:0"]}

Invoking any other model should return AccessDenied. Use that as a test of the scope. Give the credentials to the server through a role, never through a key in a file or in the web app.

What the gate does

A Strands agent has one tool, and a hook reads the stored decision. It cancels the call unless that response type is allowed for that lane. A checker then reviews words, links, length and that every cited evidence id exists. See Draft and send a comment.

Check that it worked

• Request a draft in the app on a case where drafting is allowed, and confirm it arrives.
• The spend counter stays under the cap. Past it, the next draft comes from the template.

The draft carries a badge. Model draft means the model wrote it. Offline template means no model wrote it and the wording came from a fixed template, with the note “No model wrote this.”

Next: Deploy on AWS.