The abliterated API endpointfor opencode

Set it up

1. Sign in with your email, add credit, and create a key in the dashboard. Keys start with sk-nc-.
2. Put the key in an environment variable called UNCENSORED_OPENCODE_API_KEY, so it never sits in a file you might commit.
3. Paste the block below into opencode.json at the root of your project. To have it in every project, put it in ~/.config/opencode/opencode.json instead.
4. Run opencode. It starts on GLM 5.3 Flash. Type /models to switch to any other model on the list.

# macOS and Linux
export UNCENSORED_OPENCODE_API_KEY="sk-nc-..."

# Windows PowerShell
$env:UNCENSORED_OPENCODE_API_KEY = "sk-nc-..."

{
  "$schema": "https://opencode.ai/config.json",
  "provider": {
    "uncensored-opencode": {
      "npm": "@ai-sdk/openai-compatible",
      "name": "Uncensored Opencode",
      "options": {
        "baseURL": "https://uncensoredopencode.com/v1",
        "apiKey": "{env:UNCENSORED_OPENCODE_API_KEY}"
      },
      "models": {
        "glm-5.3-flash-uncensored": {
          "name": "GLM 5.3 Flash Uncensored",
          "tool_call": true,
          "reasoning": true,
          "attachment": true,
          "limit": { "context": 1048576, "output": 32000 }
        },
        "kimi-k3.1-pro-uncensored": {
          "name": "Kimi K3.1 Pro Uncensored",
          "tool_call": true,
          "reasoning": true,
          "attachment": true,
          "limit": { "context": 1048576, "output": 32000 }
        },
        "glm-5.3-uncensored": {
          "name": "GLM 5.3 Uncensored",
          "tool_call": true,
          "reasoning": true,
          "attachment": false,
          "limit": { "context": 1048576, "output": 32000 }
        },
        "deepseek-v4.1-flash-1017-uncensored": {
          "name": "DeepSeek V4.1 Flash 1017 Uncensored",
          "tool_call": true,
          "reasoning": true,
          "attachment": true,
          "limit": { "context": 1048576, "output": 32000 }
        },
        "mimo-v2.6-flash-uncensored": {
          "name": "MiMo V2.6 Flash Uncensored",
          "tool_call": true,
          "reasoning": true,
          "attachment": true,
          "limit": { "context": 1048576, "output": 32000 }
        },
        "qwen3.8-27b-uncensored": {
          "name": "Qwen3.8 27B Uncensored",
          "tool_call": true,
          "reasoning": true,
          "attachment": true,
          "limit": { "context": 1048576, "output": 32000 }
        },
        "glm-5.3-flash-cyber": {
          "name": "GLM 5.3 Flash Cyber",
          "tool_call": true,
          "reasoning": true,
          "attachment": true,
          "limit": { "context": 1048576, "output": 32000 }
        }
      }
    }
  },
  "model": "uncensored-opencode/glm-5.3-flash-uncensored",
  "small_model": "uncensored-opencode/qwen3.8-27b-uncensored"
}

This block was loaded in opencode 1.18 before it was published here: opencode models uncensored-opencode lists every model in it. It uses the provider form of the config, which opencode has read for a long time, so it works on older installs as well as current ones.

What each line does

• npm names the adapter. @ai-sdk/openai-compatible is the one for any endpoint that speaks the OpenAI chat format, which this one does.
• baseURL is https://uncensoredopencode.com/v1. opencode adds /chat/completions to it.
• apiKey reads your key from the environment with {env:UNCENSORED_OPENCODE_API_KEY}. If you would rather keep it in a file, {file:~/.secrets/uncensored-opencode} works the same way.
• limit matters more than it looks. opencode assumes a 200,000 token context for any model it does not already know, and compacts the session early to stay under it. Setting the real figure lets it use the full million.
• attachment tells opencode whether it may send the model images. It is false on GLM 5.3, which is text only.
• small_model is the model opencode uses for cheap background jobs such as naming a session. It is set to Qwen3.8 27B, the budget model on the line.

The models

Each one has its own page with the config for that model alone and what it is good and bad at inside opencode.

Why opencode stalls on a refusal

opencode runs a loop: the model reads, calls a tool, reads the result and calls the next one. The loop only keeps going while the model keeps answering with tool calls. When a model decides a step looks risky, it answers with a paragraph about why it will not help, and that paragraph has no tool call in it. opencode treats it as the end of the turn. The files are half edited, the plan is half done, and nothing is actually wrong with opencode or with your prompt.

It happens most on work that is ordinary for developers and merely sounds alarming: a scraper, a port scanner for your own network, a proof of concept for a bug you are fixing, anything that touches credentials or malware samples. A model with the refusals removed answers the step like any other, so the loop carries on. More on this in when a model refuses in opencode.

More for opencode

Questions