How to use Kiro IDE with any OpenAI model via MITM Passthrough

The Kiro IDE desktop app is an incredible coding environment, but it has one massive limitation: you cannot configure custom API endpoints. It insists on talking directly to Kiro's cloud infrastructure.

If you want to use Kiro IDE but power it with local Ollama models, Azure OpenAI, or a Claude team subscription you already pay for, you need a MITM (Man-in-the-Middle) proxy.

Between v0.5.65 and v0.5.74, we completely overhauled kRouter's Kiro MITM intercept engine. Here is everything that changed and why it matters.

The Translator Bug (Pre-v0.5.65)

Kiro uses AWS Bedrock's backend format internally. When the Kiro IDE sends a request, it uses a complex conversationState JSON schema with Bedrock-specific conventions.

Previously, kRouter would intercept this request, translate it into the standard OpenAI format, route it to whatever provider you configured, and then try to translate the OpenAI response back into Kiro's proprietary AWS format. This 500-line translation layer was brittle. It frequently dropped images in multi-turn chats, failed with REQUEST_BODY_INVALID errors, and -- worst of all -- destroyed prompt caching.

The Tool ID Sanitization Fix (v0.5.65)

The first major fix targeted Kiro's tool calling. Kiro sends tool IDs that include characters outside the ^[a-zA-Z0-9_-]+$ regex pattern that most LLM providers enforce. When kRouter forwarded these raw tool IDs to Claude or GPT, the provider would reject the request with a schema validation error.

In v0.5.65, kRouter now sanitizes all tool IDs at the proxy boundary. It strips disallowed characters, maintains a mapping table for the duration of the conversation, and reconstructs the original IDs when passing responses back to Kiro. The IDE never knows the IDs were rewritten.

The Anti-Loop Guard (v0.5.70)

A subtle bug emerged when users ran kRouter inside the Kiro IDE's own integrated terminal. Kiro would send a request to api.kiro.dev, kRouter would intercept it, but then kRouter's own outbound request would also hit the MITM intercept -- creating an infinite loop that crashed the proxy.

The fix was the x-request-source: local header. kRouter now stamps all its own outbound requests with this header. The MITM intercept layer checks for it and passes those requests through without interception. Simple, but it eliminated a class of crash that only appeared in specific development setups.

The Persona Injection Fix (v0.5.72)

Kiro prepends its own system persona to every request. When kRouter translated the Bedrock format to OpenAI format, it would sometimes duplicate or mangle the persona block, causing the model to receive contradictory system instructions. This was especially bad for agentic workflows where precise system prompts matter.

v0.5.72 introduced persona-aware translation that correctly identifies and preserves Kiro's system block as a single system message in the OpenAI format.

First-Class Native Passthrough (v0.5.74)

The biggest change. Instead of translating at all, kRouter now supports Native Passthrough for Kiro.

If you are routing Kiro traffic to a provider that natively understands Kiro's Bedrock format (like the free Kiro OAuth provider built into kRouter), the proxy detects the user-agent: kiro header and skips the translation layer entirely.

It takes the raw byte stream from the Kiro IDE, applies RTK compression to reduce token usage, and forwards it perfectly to the destination.

Why Native Passthrough Matters for Prompt Caching

AWS Bedrock (and by extension, Kiro's backend) uses prompt caching that depends on exact byte-level prefix matching. When kRouter was translating Kiro requests to OpenAI format and back, the cached prefix was destroyed every time. Your second message in a conversation would re-process the entire context from scratch.

With native passthrough, the raw bytes are preserved. The prompt cache survives. On long conversations, this means:

• Up to 90% cache hit rates on multi-turn Kiro sessions
• Significantly lower latency on follow-up messages
• Lower token costs because cached prefixes are billed at a fraction of the standard rate

What Survives Now

With the full v0.5.74 overhaul, everything works:

• AWS Bedrock prompt caching survives intact
• Images in multi-turn chats survive
• Tool calls with non-standard IDs survive
• System personas are correctly preserved
• No infinite loops when running kRouter inside Kiro's terminal
• The 500-line translation bottleneck is bypassed entirely on native routes

How to Set It Up

1. Install kRouter: npm install -g @sifxprime/krouter
2. Start the proxy with MITM privileges: sudo krouter --mitm
3. In the kRouter dashboard, go to Settings -> Security -> MITM Intercept
4. Generate and install the root CA to your system keychain
5. Enable interception for api.kiro.dev (or the relevant Kiro domains)

Now, open the Kiro IDE. When you type a prompt, it hits your local kRouter instance. kRouter routes that traffic to whatever combo you have configured in the dashboard.

You get Kiro's UI, powered by whatever models you want, with prompt caching preserved and 20-40% fewer tokens burned thanks to RTK compression. Check the full changelog for the complete v0.5.65-v0.5.74 history.