Crush

Crush takes kiro-provider as a provider of type anthropic, so it speaks Anthropic Messages to the gateway. Both the Claude and the GPT-5.6 models work.

Add the provider

Put the provider in ~/.config/crush/crush.json, or in crush.json at the root of a project:

{
  "$schema": "https://charm.land/crush.json",
  "providers": {
    "kiro-provider": {
      "name": "kiro-provider",
      "type": "anthropic",
      "base_url": "http://127.0.0.1:8787",
      "api_key": "$KIRO_GATEWAY_API_KEY",
      "extra_headers": {
        "Authorization": "Bearer $KIRO_GATEWAY_API_KEY",
        "x-kiro-output-token-limit-mode": "advisory"
      },
      "models": [
        {
          "id": "claude-opus-5-5",
          "name": "Claude Opus 5.5",
          "context_window": 1000000,
          "default_max_tokens": 32000,
          "can_reason": true
        },
        {
          "id": "gpt-5.6-sol",
          "name": "GPT-5.6 Sol",
          "context_window": 1000000,
          "default_max_tokens": 32000,
          "can_reason": true
        }
      ]
    }
  },
  "models": {
    "large": { "model": "claude-opus-5-5", "provider": "kiro-provider" },
    "small": { "model": "claude-opus-5-5", "provider": "kiro-provider" }
  }
}

Export the key in the shell that starts Crush:

export KIRO_GATEWAY_API_KEY='<one of your api_keys>'
crush
crush run "Reply with exactly: KIRO_OK"

To work with GPT-5.6 Sol, replace the models block above with this one, which selects it with thinking on:

{
  "models": {
    "large": { "model": "gpt-5.6-sol", "provider": "kiro-provider", "think": true },
    "small": { "model": "claude-opus-5-5", "provider": "kiro-provider" }
  }
}

What the settings do

Setting	Why
base_url	Without /v1: Crush adds /v1/messages itself.
extra_headers	Authorization: Crush sends the key as x-api-key, plus an Authorization: Bearer header with no key in it. kiro-provider reads Authorization first, so without this entry it answers invalid_api_key. x-kiro-output-token-limit-mode: advisory: Crush sends max_tokens with every request, which kiro-provider refuses for GPT-5.6 without this header.
think	true for GPT-5.6 models. GPT-5.6 can send its reasoning signature after the answer has begun. With thinking on, kiro-provider waits for it; without, such a reply fails with Upstream returned invalid reasoning metadata.
context_window	Copied from context_limit in GET /v1/models.

The small model writes session titles. Keep it on a Claude model, or give a GPT-5.6 model "think": true there too.

Which models work

kiro-provider passes max_tokens on to Kiro for four Claude models: claude-opus-5-5, claude-opus-5, claude-sonnet-5 and claude-fable-5-1, from 1,024 to 128,000. With the advisory header, gpt-5.6-sol, gpt-5.6-terra and gpt-5.6-luna work too. Crush sends max_tokens with every request, so any other model is refused with an invalid_request_error about it. GET /v1/models gives the numbers for a new entry:

curl -s http://127.0.0.1:8787/v1/models -H "Authorization: Bearer $KIRO_GATEWAY_API_KEY" |
  jq -r '.data[] | [.id, .context_limit, .output_limit] | @tsv'