Use the Kiro Adapter

Use the nvidia.fabric.kiro adapter to run Kiro CLI V3 through its standard Agent Client Protocol (ACP) server. Each NeMo Fabric runtime starts one kiro-cli acp --agent-engine=v3 child process and creates one persistent ACP session. Ordered NeMo Fabric invocations reuse that session; independent runtimes use separate processes and sessions.

Install Kiro and the Adapter

Install Node.js 22.19.0 or newer, then install and authenticate Kiro CLI:

curl -fsSL https://cli.kiro.dev/install | bash
kiro-cli login
npm install nemo-fabric-adapters-kiro

For CI or another non-interactive environment, set KIRO_API_KEY instead of using browser login. The adapter uses a caller-installed Kiro CLI and never installs or updates Kiro at runtime.

For source development from the repository root:

just install-typescript-kiro
npm run build --prefix adapter-contract/typescript
npm run build --prefix adapters/typescript --workspace nemo-fabric-adapters-common
npm run build --prefix adapters/typescript --workspace nemo-fabric-adapters-kiro

Configure the Adapter

The following configuration registers the source descriptor and uses Kiro’s current default model:

from nemo_fabric import DiscoveryConfig
from nemo_fabric import EnvironmentConfig
from nemo_fabric import FabricConfig
from nemo_fabric import HarnessConfig
from nemo_fabric import MetadataConfig
from nemo_fabric import RuntimeConfig
config = FabricConfig(
    metadata=MetadataConfig(name="kiro-review"),
    discovery=DiscoveryConfig(
        local_paths=["adapters/typescript/kiro/kiro.fabric-adapter.json"]
    ),
    harness=HarnessConfig(
        adapter_id="nvidia.fabric.kiro",
        settings={"permission_mode": "allow_once"},
    ),
    runtime=RuntimeConfig(input_schema="text", output_schema="message"),
    environment=EnvironmentConfig(
        provider="local",
        workspace=".",
    ),
)

Use the packaged descriptor at node_modules/nemo-fabric-adapters-kiro/kiro.fabric-adapter.json after an npm install. When a model is supplied, the adapter verifies it against the session’s advertised model option before the first invocation.

Adapter Settings

The settings schema is closed:

The adapter never selects a persistent allow option. permission_mode controls ACP permission responses; it is not a normalized NeMo Fabric tool allowlist.

MCP Servers

The adapter sends normalized MCP definitions as client-owned inputs to session/new:

• Stdio servers support args and env, require an absolute executable path, and do not accept HTTP headers.
• Streamable HTTP and SSE servers support custom_headers, but do not accept process arguments or environment variables.
• Network endpoints require HTTPS. Loopback HTTP remains available for local development.
• Authentication objects, per-server tool filters, and extensions are rejected rather than ignored.

Kiro’s initialization response is authoritative. Startup fails when the running Kiro deployment does not advertise the configured network transport. When MCP servers are configured, startup also waits for Kiro to report every server as connected before accepting the first invocation.

Supported Surface

The initial adapter accepts models and mcp. Kiro model, mode, and effort selection use advertised ACP session configuration. Successful assistant text, terminal stop reasons, session identifiers, and reported token usage are normalized into NeMo Fabric results.

NeMo Fabric system instructions, skill paths, built-in tool policy, arbitrary model endpoints, Relay telemetry, and streaming are not advertised. Kiro may support related native features, but this adapter does not claim them until NeMo Fabric can configure and verify them through a stable Kiro interface.

Try the Code-Review Example

.venv/bin/python -m examples.code_review_agent --variant kiro --plan
.venv/bin/python -m examples.code_review_agent \
  --variant kiro \
  --input "Review calculator.py" \
  --show-output

Set KIRO_API_KEY before these commands only in a non-interactive environment.

The example intentionally does not project the shared NeMo Fabric skill or system instruction because the initial Kiro adapter does not advertise those fields.

Runtime health

Fabric.check() supports live. The adapter observes the selected ACP child process identifier, exit status, and signal status through the shared process runner. A retained session identifier is insufficient. A busy process remains live, and a stopped or exited process fails the liveness observation.

active and ready are unsupported. Checks do not send prompts, change session settings, or cancel invocations.

Refer to the SDK health-check overview for shared behavior and a comparison of native evidence across adapters.