Uncensored coding AI in Claude Code, Cursor and Open WebUI

What an uncensored coding AI changes

A coding assistant refuses in two places. The first is the obvious one: a request for exploit code, a packer, a credential harvester for a lab, a deobfuscated phishing kit. The second is quieter: the model narrows the answer, adds a disclaimer where a function should be, or stops the agent loop with a question. In an agent either kind is a dead end, because a retry against a model trained to refuse returns the same answer at the same price.

An uncensored code build has that refusal removed from the weights, by abliteration or by a fine-tune on data with no refusals in it (the two edits). It reads a decompiled binary, a packet capture or a malware sample as text to analyse and writes the code asked for. It is not a different kind of programmer: the base model's skill at Python, Rust or SQL is what it was.

Three code builds are on the key today, listed with the general assistant for comparison. Prices are per million tokens from the live catalog.

• uncensored-coding-agent: a UncensoredLLM house build for agent loops, with 1,048,576 tokens of context, which is a whole repository in one request.
• glm-5.3-flash-code-uncensored: the uncensored code build of Z.ai's GLM-5.3 Flash (320B total, 18B active, MIT, card read 2026-10-26).
• uncensored-code-assistant: the UncensoredLLM house build for one-shot answers in a chat panel, the highest-priced row, with 131,072 tokens of context.

The general assistant glm-5.3-uncensored handles most coding too; the code builds earn their place when the task is security work or a long agent session.

Claude Code: through the router

Claude Code sends Anthropic-format requests and every build here answers the OpenAI format, so claude-code-router sits between them. The Claude Code guide covers the install and the check request; this is the provider block with the code builds in it:

APIKEY is any string you choose and is what Claude Code sends the router; api_key is your UncensoredLLM key. Switching builds mid-session is /model with another id from the provider block. The router's README is at musistudio/claude-code-router.

Cursor: Models settings

Cursor takes a custom OpenAI key and base URL under Settings, Models. Three settings and the id, and the build answers in the chat panel and in Agent mode. Tab completions stay on Cursor's own models.

OpenAI API Key

sk-ullm-live-…

your UncensoredLLM key

Override OpenAI Base URL

https://api.uncensoredllm.ai/v1

on; the /v1 at the end matters

Model

uncensored-coding-agent

added by hand under the model list, typed exactly

Field by field, with the errors and fixes, in the Cursor guide.

Open WebUI: one connection, a preset per build

Under Admin, Connections, add an OpenAI API connection with URL https://api.uncensoredllm.ai/v1 and the key. The model list loads from the API; keep the picker short by allowing only the code builds, and give each one a Workspace preset with the system prompt your team uses. The Open WebUI guide has the screens, the context-window setting and a costed chat.

What a coding day costs

Computed at render from the catalog of 2026-10-27: 150 agent turns, each with 60,000 tokens of cached prefix (the repository and the conversation so far), 2,500 tokens of new input and 1,200 tokens of edits out. Cached tokens bill at the build's cache-read rate where the catalog has one.

Turns × (cached × cache-read price + input × input price + output × output price), catalog of 2026-10-27. The check column is the daily real-model check.

The cached prefix dominates an agent day, so a build with a low cache-read rate is cheap to keep in a long session even when its output price is not the lowest. Every rate is on the pricing page; the balance is prepaid and each request settles at these rates.

Check the build before the agent

One request from a terminal shows the build answers the kind of question a stock assistant declines, and prints the cost header:

A 200 with x-uplink-charge is the whole test. A 401 is the key; a 404 with model_not_found is the id, usually a display name typed where the id goes. Create the key in the console and set a spend limit on it before pasting it into a shared tool.

Security work: what changes, and what stays forbidden

What changes: the model writes the exploit for the CVE you are patching, walks through the malware sample you are classifying, rewrites a payload so a detection rule can be tested, and reads a phishing kit as source. The red-teaming page lists the security builds tuned on security corpora and the settings for the agent tools; for a coding agent the code builds above do the same job with more code and less commentary.

What stays forbidden is on the acceptable-use page and applies to every key: accessing, attacking or testing systems, networks or accounts you do not own or have written permission to test; building or spreading malware against others; fraud, phishing and impersonation. Security and red-team builds are for assessing systems you own or are authorized in writing to test, and the policy says to keep proof of that authorization. Chat requests are not filtered for content by the gateway; the model answers as its build answers, and the policy covers what you send. A key that breaks it is revoked, and the API answers 401 within 60 seconds.

Questions