A hard dollar cap per Kilo CLI session with your own API key (x-session-id, config only)

[domondi1]

If you use the Kilo CLI with your own API key (a custom OpenAI-compatible provider rather than Kilo's gateway), you can give every session its own hard dollar ceiling with config only. The CLI already sends its session id as x-session-id on every request to a custom provider, so a local gateway can use that as the run id and refuse a call with HTTP 402 before it reaches the provider once the session's budget is used. Disclosure: I maintain the gateway (Inferrail, open source).

pip install inferrail                      # 0.4.16+
export OPENAI_API_KEY=sk-...               # stays in the gateway
inferrail serve --quickstart --app-mode --work-id-header x-session-id

~/.config/kilo/kilo.json:

{
  "provider": {
    "capped": {
      "npm": "@ai-sdk/openai-compatible",
      "options": {
        "baseURL": "http://127.0.0.1:8000/v1",
        "apiKey": "unused",
        "headers": { "X-Inferrail-Budget-Usd": "3.00" }
      },
      "models": { "gpt-4o-mini": { "name": "gpt-4o-mini" } }
    }
  },
  "model": "capped/gpt-4o-mini"
}

Each session (ses_...) gets its own $3, and inferrail report --by work_id lists calls, tokens and dollars per session. For a CI or scripted run that should be one budget across several sessions, add "X-Inferrail-Attribute-Work-Id": "<run id>" to the headers; an explicit id wins.

What I checked (@kilocode/cli 7.8.3, inferrail 0.4.16, kilo run against a stand-in OpenAI endpoint with fixed token counts): two sessions recorded under their own ids; with a budget too small for the next call, kilo run printed Error: budget 'work_id:ses_...' would be exceeded after one refused request, nothing reached the upstream, no retry loop.

Caveats: each call reserves its worst case from the max output tokens up front, so budgets need headroom (about $0.02 per call for gpt-4o-mini at the CLI's default here). A dollar budget needs a price for the model on the gateway side (most OpenAI and Anthropic models are built in, inferrail models). This is for your own API keys, not Kilo's hosted gateway, which has its own billing. The gateway stores usage and cost per call, never prompts or code.