Kilo Code: custom provider and kilo.jsonc limits

Kilo Code version 7 is built on the opencode core and shares one config file across VS Code, JetBrains and the terminal. You add uttapen as a Custom provider. This page was checked against Kilo's official docs for 7.x.

Adding the provider in the UI

1. Settings (gear) → Providers tab → Custom provider at the bottom of the list.
2. Fill in the dialog:

3. Once the URL is valid, Kilo fetches the catalogue from GET /v1/models; pick the models you want from that list and press Submit.

Output and context limits in kilo.jsonc

The dialog has no token-limit fields; those live in the config file, either ~/.config/kilo/kilo.jsonc globally or kilo.jsonc in the project root.

{
  "$schema": "https://app.kilo.ai/config.json",
  "provider": {
    "uttapen": {
      "npm": "@ai-sdk/openai-compatible",
      "options": {
        "apiKey": "{env:UTTAPEN_API_KEY}",
        "baseURL": "https://api.uttapen.ir/v1"
      },
      "models": {
        "anthropic/claude-sonnet-5": {
          "name": "Claude Sonnet 5",
          "tool_call": true,
          "limit": { "context": 200000, "output": 8192 }
        },
        "openai/gpt-5-nano": {
          "name": "GPT-5 nano",
          "limit": { "context": 400000, "output": 1024 }
        }
      }
    }
  },
  "model": "uttapen/anthropic/claude-sonnet-5",
  "small_model": "uttapen/openai/gpt-5-nano"
}

• limit.output is sent to us as max_tokens, and Kilo's internal default for it is 32,000. Since uttapen reserves the cost of that ceiling before the request goes out (pricing), a large number locks up your balance for nothing. 8192 is enough for code edits.
• Leave limit out and both values resolve to zero, which disables automatic context management: the conversation is never compacted and each prompt grows larger and pricier.
• {env:UTTAPEN_API_KEY} reads the key from the environment, so it never sits in the file.
• Kilo model names are provider_id/model_id and only the first / separates them, so uttapen/anthropic/claude-sonnet-5 is correct.

Always set a Small Model

Kilo uses the Small Model (the Settings → Models tab, or the small_model key) for session titles, commit messages and prompt enhancement. Leave it empty and Kilo first looks for a cheap model on the same provider; because our custom provider isn't in its built-in catalogue, it either falls back to Kilo's own service or reuses your main model with the same large reservation. Point it at something cheap such as openai/gpt-5-nano or google/gemini-2.5-flash-lite.

For the main model, take ids from the model list: anthropic/claude-sonnet-5 and openai/gpt-5 for heavy work, qwen/qwen3-coder and deepseek/deepseek-v3.2 when cost matters.

Checking that it works

Send a small request from the Kilo panel or the CLI, then look at Dashboard → Usage: you should see two rows, one for the main model and a tiny one for the Small Model.

Troubleshooting

If your available balance can't cover the full max_tokens, uttapen trims the output ceiling to fit the balance (never below 1,024 tokens) and returns the granted value in the X-Uttapen-Max-Tokens header — which is when you see Kilo's "response may be incomplete" warning. Only if even 1,024 tokens don't fit do you get 402.

Other tools live in integrations; error codes are in errors.