What you need
- the Codex CLI installed.
- An XField API key in the environment variable
XFIELD_API_KEY. No key yet? Request access. - The base URL
https://xfield.cloud.brutal.ai/api/v1. It already ends in/v1.
Setup
Codex talks to models through the OpenAI Responses API, which XField serves at https://xfield.cloud.brutal.ai/api/v1/responses. Add to ~/.codex/config.toml:
model = "deepseek-v4.1-flash-1017-uncensored"
model_provider = "field"
model_reasoning_effort = "high"
model_context_window = 1000000
[model_providers.field]
name = "XField"
base_url = "https://xfield.cloud.brutal.ai/api/v1"
env_key = "XFIELD_API_KEY"
wire_api = "responses"Run codex, then /status to confirm model and provider. Switch per run with -m or any key with -c:
codex -m glm-5.3-uncensored
codex -c model_reasoning_effort="max" "Review this diff for auth bypasses"Which model to pick
| Job | Model id | Why |
|---|---|---|
| Most coding and agent work | deepseek-v4.1-flash-1017-uncensored | Leads our models on coding and security benchmarks, fast, cheap once the prompt is cached |
| Hard agentic coding, very long outputs | kimi-k3.1-pro-uncensored | Moonshot AI's flagship, up to 922K output tokens; costs more, so use it where it matters |
| Long reports, a second opinion | glm-5.3-uncensored | 128K output, a different model family |
| Cheap background work, summaries | glm-5.3-flash-uncensored | The lighter GLM, built for volume |
| Quick chat and drafts | qwen-3.8-27b-uncensored | Smallest and fastest to answer |
Context and output limits
Tell the Codex CLI the limits, otherwise it guesses low and condenses too early. Turn image input on only for models that read images.
| Model id | Context | Max output | Reads images |
|---|---|---|---|
deepseek-v4.1-flash-1017-uncensored | 1M | 64K | Yes |
kimi-k3.1-pro-uncensored | 1M | 922K | Yes (and video) |
glm-5.3-uncensored | 1M | 128K | No |
glm-5.3-flash-uncensored | 1M | 32K | Yes |
qwen-3.8-27b-uncensored | 128K | 32K | Yes |
Check the key and URL first
If this returns a JSON list of models, the base URL and key are right and any remaining problem is in the tool's settings:
curl -sS "https://xfield.cloud.brutal.ai/api/v1/models" -H "Authorization: Bearer $XFIELD_API_KEY"Troubleshooting
- 401 Unauthorized. XFIELD_API_KEY is not set in the shell that started Codex.
- Requests go to the wrong path. base_url must end in /v1; Codex appends /responses itself.
- Model not found. The id must match exactly, including
-uncensored. List them with the call above. - The tool condenses or cuts off too early. It does not know these models. Set context and output from the limits table.
Why an uncensored model in a coding tool
Stock models stop mid-task on legitimate work: an exploit proof of concept for your own bug bounty, a fuzzing harness, malware triage, a detection rule that has to describe the attack it detects. In an agent loop one refusal breaks the whole run. Our models are abliterated: the direction in the weights that leads to refusals is removed, the rest of the model stays the same. How abliteration works.
More setups
- Docs: Codex (profiles, reasoning effort)
- Use Claude Code with an uncensored model
- Uncensored LLM APIs compared (2026)
- Other OpenAI-compatible tools
Get a key
XField is the model API of Brutal AI in Vienna. It is hosted in the EU, does not store prompts or completions (data handling), and access is reviewed company by company under our acceptable use policy. Request a key.