opencode-kiro-provider

Use every model your Kiro subscription offers inside OpenCode: GPT‑5.6 Sol/Terra/Luna, Claude Opus 5.5, Sonnet 5.5, Haiku, the open-weight models, and Kiro's Auto router.

OpenCode stays in charge of the agent loop (its tools, permissions, agents and UI), and Kiro provides the models.

• Always the current model list. The plugin asks Kiro which models your account can use (ListAvailableModels, the same call the Kiro IDE makes). When Kiro adds a model, it shows up in OpenCode on its own, with no plugin update or config edit.
• GPT and Claude, both done right. OpenCode's reasoning variants (--variant high, or the variant picker) are sent the way each model family expects: reasoning.effort for GPT, output_config.effort for Claude. The allowed levels come from Kiro, per model.
• Real limits and costs. Each model gets its real context window from Kiro, and its credit multiplier appears in the model name, e.g. Claude Opus 5 (2.2x).
• Real token usage, including cache reads, reported to OpenCode.
• No extra login. It reuses your kiro-cli login: IAM Identity Center (organization SSO), Builder ID, or an external IdP. Tokens are refreshed automatically.
• Full agent support: streaming, tool calls (including parallel ones), images and reasoning output.

Unofficial project. It isn't affiliated with or endorsed by Amazon Web Services or Kiro. It calls the same Kiro service APIs as the official clients, using your own Kiro login, so your use is subject to your Kiro/AWS terms.

Install

Requirements: OpenCode ≥ 1.14 and kiro-cli.

1. Log in to Kiro once:

   kiro-cli login

2. Add the plugin to ~/.config/opencode/opencode.json:

   {
     "$schema": "https://opencode.ai/config.json",
     "plugin": ["[email protected]"]
   }

   OpenCode installs it from npm on the next start. Remove any other Kiro plugin (such as opencode-kiro-auth) from the list, because both register the provider id kiro.
3. Start OpenCode, type /models, and pick a model under Kiro.

If no Kiro models show up on the very first start, restart OpenCode once, or type /connect in OpenCode, choose Kiro, then choose Use my kiro-cli login.

Updating: OpenCode installs the package once and reuses its cached copy. Pin a version as shown above, and change it to upgrade. Or delete the cached package under ~/.cache/opencode/packages/ and restart OpenCode. The model list doesn't depend on the plugin version, because it's fetched from Kiro.

Setting this up with an AI agent? Point it at INSTALL-AGENT.md. It's a step-by-step guide with checks and stop points.

From source

git clone https://github.com/JarbesGoldoni/opencode-kiro-provider.git

Then put the clone's absolute path in the plugin list instead of the package name, e.g. "/home/you/opencode-kiro-provider". No build step is needed, because OpenCode runs the TypeScript directly.

Usage

Everything happens inside OpenCode, just like with its built-in providers:

• Pick a model: type /models and choose one under Kiro, e.g. GPT-5.6 Sol (4.4x) or Claude Opus 5 (2.2x). The number is Kiro's credit multiplier.
• Set reasoning effort: press ctrl+t to cycle through the levels the model supports (for example low → high → max), or type /variants to pick one. Models without effort control have no variants.
• Work as usual: chat, edit files, run commands and use your agents and MCP servers. OpenCode runs the tools, and Kiro provides the model.

Model availability depends on your Kiro plan and region. Some models (for example GPT‑5.6) are only offered in some regions. /models shows exactly what Kiro reports for your account.

The Kiro models work in OpenCode's non-interactive mode too:

opencode run -m kiro/gpt-5.6-sol --variant high "refactor src/app.ts"

Configuration

All environment variables are optional:

| Variable | Purpose | |---|---| | KIRO_DEBUG=1 | Log requests and responses (never tokens) to ~/.cache/opencode-kiro-provider/debug.log | | KIRO_DEBUG_BODY=1 | Also log full request bodies (these include your prompts) | | KIRO_REFRESH_MODELS=1 | Ignore the 6-hour model-list cache and refetch | | KIRO_AUTH_KIND | Force a login type: idc (IAM Identity Center), social (Builder ID) or external-idp | | KIRO_HTTP_USER_AGENT | Override the HTTP user-agent header | | KIRO_USER_AGENT | Override the x-amz-user-agent header | | KIROCLI_DB_PATH | Non-standard kiro-cli database location |

To change a model's settings, define it under provider.kiro.models in opencode.json. Your entry replaces the plugin's entry for that model id.

Troubleshooting

Errors quote Kiro's own message, followed by a hint:

| Message | Meaning | |---|---| | No Kiro login found | Run kiro-cli login. | | "…login has expired or was revoked" | Run kiro-cli login again. | | "…rejected this client or model" | Your login is fine, but Kiro refused the call, usually because that model isn't enabled for your account or region. Rerun with KIRO_DEBUG=1 and check the log. | | A new Kiro model is missing from /models | The list is cached for 6 hours. Start OpenCode with KIRO_REFRESH_MODELS=1 to refetch it. |

How it works

OpenCode ──OpenAI chat format──▶ plugin fetch ──GenerateAssistantResponse──▶ runtime.<region>.kiro.dev
   ▲                                 │                                              │
   └────── OpenAI SSE chunks ◀───────┴────────── AWS event-stream ◀─────────────────┘

The provider registers as @ai-sdk/openai-compatible with a custom fetch that:

1. converts OpenAI messages to Kiro's conversationState: history alternates user/assistant turns, tool results are paired with their tool uses, and tool names and JSON schemas are normalized the way Kiro expects,
2. sends the request to runtime.<region>.kiro.dev, falling back to the legacy q.<region>.amazonaws.com host,
3. decodes Kiro's binary event stream (text, reasoning, tool calls, token usage) back into OpenAI stream chunks.

OpenCode gives GPT models its apply_patch tool instead of edit/write. That's OpenCode's own behavior, and it works through this plugin.

Development

bun install
bun test           # unit tests
bun run typecheck
bun run e2e        # real opencode binary against a mock Kiro server

bun run e2e uses throwaway config folders and a fake kiro-cli login, so it never touches your real setup. The mock server enforces Kiro's history, tool-pairing and header rules.

To check against your real Kiro account, bun run smoke lists your models and sends one short prompt to gpt-5.6-sol and claude-opus-5 (a few credits). Pass model ids to test others: bun run smoke gpt-5.6-terra auto.

Issues and pull requests are welcome.

License

MIT