Run uncensored models in the Codex CLI with one config block, and in Claude Code through a LiteLLM proxy. Which uncensored models finish agent tasks, and what they cost in a coding agent.

Claude Code and the Codex CLI are built for their makers' models, but both can run another model. This is how to run an uncensored one in each, with OpenUncensored as the backend: one key, 61 uncensored models, OpenAI compatible. There is also a third option for Claude Code: keep Claude as the agent and give it uncensored models as a tool.

Buy a prepaid key at openuncensored.com/credits: no account, no sign up, just a card and a key (ou-...). Or hand your agent openuncensored.com/agents.md and let it get one for you.

Codex CLI (direct)

Codex speaks the OpenAI Responses API, which OpenUncensored serves at https://api.openuncensored.com/v1/responses. Add a provider to ~/.codex/config.toml:

[model_providers.openuncensored]
name = "OpenUncensored"
base_url = "https://api.openuncensored.com/v1"
env_key = "OU_API_KEY"
wire_api = "responses"

[profiles.uncensored]
model_provider = "openuncensored"
model = "glm-5.3-flash-apex-abliterated"

Then export OU_API_KEY=ou-... and run codex --profile uncensored. File edits, shell commands and multi-step tasks work as usual.

For harder tasks, add a profile with a reasoning model and an effort level:

[profiles.uncensored-thinking]
model_provider = "openuncensored"
model = "deepseek-v4.1-flash-1017-uncensored-thinking"
model_reasoning_effort = "high"

Claude Code (through a local proxy)

Claude Code talks the Anthropic Messages API, and OpenUncensored is OpenAI compatible, so put a translator in between. LiteLLM's proxy does this and is the gateway Claude Code's own docs describe:

# litellm.yaml
model_list:
  - model_name: uncensored
    litellm_params:
      model: openai/glm-5.3-flash-apex-abliterated
      api_base: https://api.openuncensored.com/v1
      api_key: os.environ/OU_API_KEY
  - model_name: uncensored-alt
    litellm_params:
      model: openai/glm-5.3-flash-uncensored
      api_base: https://api.openuncensored.com/v1
      api_key: os.environ/OU_API_KEY

pip install 'litellm[proxy]'
export OU_API_KEY=ou-...
litellm --config litellm.yaml --port 4000

# in another terminal
export ANTHROPIC_BASE_URL=http://localhost:4000
export ANTHROPIC_AUTH_TOKEN=anything
claude --model uncensored

Switch models inside a session with /model uncensored-alt. The proxy only listens on your machine; the key stays in its environment.

Claude Code with uncensored models as a tool

If you want Claude to stay the agent and only hand some requests to an uncensored model, add our MCP server instead:

claude mcp add --transport http open-uncensored https://openuncensored.com/mcp \
  --header "Authorization: Bearer ou-..."

Claude then has tools to list models, read the rankings, look up a model, check your balance and send a prompt to any uncensored model, billed to your key. The agent skill teaches Claude Code, Codex, Cursor and other agents to pick a model, get a key and call the API: npx skills add DhruvAtreja/open-uncensored-skills.

Which model to use in an agent

Coding agents need a model that finishes multi-step tool tasks without looping, not just one that answers. These are the current best uncensored models on our agent tests:

Model	id	Agent tasks	Loops	$ per 1M in + out
Apex Flash 1 Abliterated	`glm-5.3-flash-apex-abliterated`	3/5	0%	$0.650
GLM 5.3 Flash Uncensored	`glm-5.3-flash-uncensored`	3/5	40%	$0.620
Gemma 4 31B Gembrain Uncensored Heretic	`gemma-4-31b-gembrain-uncensored-heretic`	3/5	40%	$0.612
DeepSeek V4.1 Flash 1017 Uncensored	`deepseek-v4.1-flash-1017-uncensored`	2/5	0%	$1.65
MiMo V2.6 Flash Uncensored	`mimo-v2.6-flash-uncensored`	2/5	0%	$1.96
DeepSeek V4.1 Flash 0910 Uncensored	`deepseek-v4.1-flash-0910-uncensored`	2/5	20%	$1.65

The five tasks:

File search & edit (3 tools, up to 30 turns): Find a config value across a fake repo and patch it.
Web research (2 tools, up to 26 turns): Search, open pages, reconcile two conflicting sources.
Multi step math (1 tools, up to 30 turns): Chain eight calculator calls without losing state.
Code debugging (3 tools, up to 28 turns): Run tests, read a traceback, fix, rerun until green.
Data pipeline (5 tools, up to 24 turns): Load, filter, join and export a CSV through five tools.

What it costs

Coding agents resend the whole conversation at every step, so most of their tokens are cached input. The cached price on each model page decides the bill more than the headline price. A prepaid key also caps spending: when the credit runs out, requests stop with 402 insufficient_quota instead of running up a bill.

Troubleshooting

Status	Code	What it means	What to do
401	`invalid_api_key`	Missing or wrong key	Send `Authorization: Bearer ou-...`
402	`insufficient_quota`	The key's credit is used up	Top up the same key at /credits
404	`model_not_found`	No model with that id	Copy the id from /models or `GET https://api.openuncensored.com/v1/models`
400	`invalid_routing_options`	`only`/`ignore` exclude every provider, or an unknown strategy	Loosen the routing block
4xx	`upstream_rejected_request`	The provider rejected the request itself	Read the message: usually a parameter that model does not support
429	`rate_limit_exceeded`	Every provider for the model is rate limited	Retry after the `Retry-After` header
502	`all_providers_failed`	Every provider was tried and failed	Retry, or pick another model

The agent loops or repeats a tool call. Pick a model with a low loop rate from the table above.
Tool calls arrive as plain text. The model is not following the tool format. Try another model; agent scores show which ones handle tools.
Claude Code says the model does not exist. The name after --model must match a model_name in litellm.yaml, not our model id.

Setup for Cursor, Cline, Kilo, Roo and OpenCode is in this guide.

How to use an uncensored model in Claude Code or the Codex CLI