Eligible new accounts get $0.10 in free API credit. Once per account; abuse limits apply. Create an account

Outrun
Sign in Get your key

Why coding agents refuse when abliterated models don't

You point a coding agent at an abliterated model and it still says no. The refusal usually comes from the harness around the model, not the model itself.

ExplainerBy Outrun2 min read

Short answer: the app is refusing, not the model. Abliteration removes the model's reflex to refuse, but the model still follows instructions, and most coding agents send instructions that tell it to decline.

The four usual causes

  1. The app's built-in prompt says to refuse. Claude Code, Cline, Roo Code and others send thousands of tokens of their own rules with every request. The model obeys them.
  2. A second model handles part of the session. Titles, summaries, sub-agents and command checks often go to a separate "small" or "fast" model that you never changed.
  3. The app blocks the request itself. Some apps check content, or send it through their own servers, before it reaches your API.
  4. An old refusal is still in the chat. One refusal in the history makes the next one more likely.

Find which one it is

  • No request in your Outrun Activity? The app blocked it (cause 3).
  • A different model ID in Activity? A side slot is still on the default (cause 2).
  • Works with a plain API call? It's the app's prompt (cause 1). Test with:
curl https://api.outrun.to/v1/chat/completions \
  -H "Authorization: Bearer $OUTRUN_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"glm-5.3-abliterated","max_tokens":400,
       "messages":[{"role":"user","content":"YOUR REQUEST"}]}'

Fix it

  • For coding, use Zoo Code or Roo Code. Custom modes let you change the built-in instructions, and GLM 5.3 Abliterated works with Roo Code's tool calls. Zoo Code is the maintained fork and takes the same Outrun settings. Setup guide.
  • For chat, use Outrun Chat. Pick an abliterated model and talk to it directly, with no coding-agent rules in the way.
  • Start a new chat after a refusal. Don't argue in the same one.

All options: uncensored AI API.

Abliterated models reduce refusals; replies can still vary. They are for adults, and Outrun's Terms apply to every request.

← All posts