Connect Kilo Code to Token Factory

​

Prerequisites

• A Token Factory API key. See Authentication.
• Kilo Code 7.8.8 or later: the CLI (npm install -g @kilocode/cli) or the VS Code or JetBrains extension.
• A model from the Token Factory catalog.

​

Context window and compaction

• limit.context is the model’s context_window.
• limit.output is the model’s max_output_tokens. Kilo Code sends max_tokens = the smaller of this and its built-in 32,000-token cap, reduced only when the context is nearly full.
• limit.input sets where Kilo Code compacts: it summarizes the conversation once a request reaches limit.input minus 20,000 tokens (its default compaction.reserved while the output cap is at least 20,000). Without limit.input, Kilo Code compacts only at limit.context − max_tokens, with no headroom.

export CORVEX_API_KEY="sk-corvex-..."

{
  "provider": {
    "corvex": {
      "npm": "@ai-sdk/openai-compatible",
      "name": "Corvex Token Factory",
      "options": {
        "baseURL": "https://api.tokenfactory.corvex.cloud/v1",
        "apiKey": "{env:CORVEX_API_KEY}"
      },
      "models": {
        "zai-org/GLM-5.3": {
          "name": "GLM 5.3",
          "reasoning": true,
          "tool_call": true,
          "limit": { "context": 393216, "input": 334100, "output": 393216 }
        },
        "deepseek-ai/DeepSeek-V4-Flash-0731": {
          "name": "DeepSeek V4 Flash 0731",
          "reasoning": true,
          "tool_call": true,
          "limit": { "context": 1048576, "input": 903979, "output": 1048576 }
        }
      }
    }
  },
  "model": "corvex/zai-org/GLM-5.3"
}

kilo run -m corvex/deepseek-ai/DeepSeek-V4-Flash-0731 "Summarize this repository."

​

Compatibility notes

• Model IDs. Use the exact id from GET /v1/models. IDs are case-sensitive and use namespace/model format; add the corvex/ provider prefix only when you select a model on the command line.
• Output budget. Reasoning tokens count toward the 32,000-token budget. The KILO_EXPERIMENTAL_OUTPUT_TOKEN_MAX environment variable changes the cap; recompute the compaction window if you change it.
• Context length. If a long task still exceeds the model’s window, the gateway returns context_length_exceeded. Start a new task or shorten the context. See Errors.
• Debug output. kilo debug config prints the resolved API key in plain text. Do not paste its output into issues or logs.

​

Related documentation

• Connect OpenCode — the same configuration format in OpenCode.
• Use the OpenAI SDK — the same endpoint from Python or TypeScript.
• Errors — authentication, rate-limit, and validation responses.