Coding tools

Kilo Code

Add Deference to Kilo Code as an OpenAI-compatible custom provider.

Kilo Code reads one config file in the CLI, VS Code and JetBrains. A custom provider takes a base URL and a key. This guide uses Chat Completions with https://deference.si/v1.

Set up

Create a key in API keys and export it as DEFERENCE_API_KEY.
Add the provider to your global config, ~/.config/kilo/kilo.jsonc (on Windows, C:\Users\<you>\.config\kilo\kilo.jsonc).

{
  "$schema": "https://app.kilo.ai/config.json",
  "model": "deference/claude-sonnet-5.5",
  "provider": {
    "deference": {
      "npm": "@ai-sdk/openai-compatible",
      "options": {
        "baseURL": "https://deference.si/v1",
        "apiKey": "{env:DEFERENCE_API_KEY}"
      },
      "models": {
        "claude-sonnet-5.5": {
          "id": "anthropic/claude-sonnet-5.5",
          "name": "Claude Sonnet 5.5",
          "tool_call": true,
          "limit": { "context": 200000, "output": 32000 }
        }
      }
    }
  }
}

Restart Kilo Code.

{env:DEFERENCE_API_KEY} works only in the global config. In a project file that you commit, Kilo ignores it. Keep the key out of project files.

The key under models is the name Kilo shows and uses. id is what Kilo sends to Deference. Put the slash-free name in model and in -m.

To add the provider in VS Code instead, open Settings, then Providers, scroll to Custom provider, and set Provider API to OpenAI Compatible with the base URL and key. Token limits and tool calling still need the file.

Verify

Run kilo models deference. The model you declared is listed.
Run kilo roll-call "^deference/". It sends a short prompt and prints the latency per model.
Run kilo run -m deference/claude-sonnet-5.5 "Reply with exactly: gateway-ok".
Open Activity. The request is the top row, with the model you chose.

Pick a model

limit.context decides when Kilo compacts the conversation. With no limit, it never does. Set both context and output yourself.
limit.output is sent as max_tokens. Kilo caps it at 32,000 by default.
tool_call: true turns on file and terminal tools.
Use @ai-sdk/openai as npm for models served through POST /v1/responses, with the same base URL.

Troubleshooting

You see	Cause	Fix
`401 invalid_api_key`	`{env:DEFERENCE_API_KEY}` sits in a project file, or the variable is not set	Move the provider to the global config and export the variable
The model does not appear	The file is not valid JSONC, or `model` does not match a key	Check commas, and use `deference/` plus the key under `models`
The agent cannot edit files	`tool_call` is off	Set `"tool_call": true`
The conversation never compacts	`limit.context` is missing	Set `limit.context`
`404 model_not_found`	`id` differs from the catalog	Copy the id from the Models page
`402 insufficient_credit`	No available credit	Add credit
`402 free_credit_not_eligible`	The account has only free credit, and this model or request does not qualify	Use an open-weight text model, set a lower `max_tokens`, or add credit
`402 key_limit_reached`	The key reached the credit limit set on it	Raise the key's limit in API keys, or use another key