---
title: "Uncensored coding AI for pentesters: Claude Code and Cursor"
canonical: https://pentestllm.ai/guides/uncensored-coding-ai
generated: 2026-11-18T03:35:59.718Z
source: https://pentestllm.ai/guides/uncensored-coding-ai
base_url: https://api.pentestllm.ai/v1
---

# Set up an uncensored coding AI for security work in Claude Code, Cursor and Open WebUI

An uncensored coding AI inside the agent a tester already runs: the code builds with the refusal trained out, wired into Claude Code via a router, Cursor and Open WebUI, one agent day priced from the live catalog, and the line the acceptable-use page draws.

Steps checked October 27, 2026 · prices as of November 18, 2026

An uncensored coding AI is a code model whose trained refusal has been edited out, running in the agent you already open every morning. This guide wires the code builds on this key into Claude Code (through a router), Cursor and Open WebUI, prices one agent day at live rates, and spells out what the edit changes for a tester and where the acceptable-use page still draws the line. The [uncensored coding AI page](https://pentestllm.ai/uncensored-coding-ai) has the short version.

Every client needs the same three things: the base URL `https://api.pentestllm.ai/v1`, a key from the [console](https://app.pentestllm.ai/console), and a model id. The [docs page](https://pentestllm.ai/docs) lists them for each other tool.

## What you need

The same three values in every tool: base URL `https://api.pentestllm.ai/v1`, a key from the [console](https://app.pentestllm.ai/console), and a model id. Put a spend limit on the key before it goes into a tool other people on the team share. To try a build with nothing installed, the prompt box on the home page gives a few free turns and [the playground](https://pentestllm.ai/playground) runs every id in the browser.

Three code builds are on the key as of this render, with the general assistant beside them for comparison. Prices are per million tokens, straight from the catalog.

| Model | Build | Input $/M | Output $/M |
| --- | --- | --- | --- |
| [Uncensored Coding Agent](https://pentestllm.ai/models/uncensored-coding-agent) uncensored-coding-agent | House build | $0.90 | $3.00 |
| [GLM-5.3 Flash Code Uncensored](https://pentestllm.ai/models/glm-5.3-flash-code-uncensored) glm-5.3-flash-code-uncensored | Uncensored | $6.00 | $6.00 |
| [Uncensored Code Assistant](https://pentestllm.ai/models/uncensored-code-assistant) uncensored-code-assistant | House build | $11.70 | $15.60 |
| [GLM-5.3 Uncensored](https://pentestllm.ai/models/glm-5.3-uncensored) glm-5.3-uncensored | Uncensored | $0.282 | $0.846 |

Prices per million tokens, read from the live catalog on 2026-11-18.

- [`uncensored-coding-agent`](https://pentestllm.ai/models/uncensored-coding-agent): the PentestLLM house build meant for agent loops. Its 1,048,576 tokens of context take an entire repository, or a decompiled binary and its strings, in one request.
- [`glm-5.3-flash-code-uncensored`](https://pentestllm.ai/models/glm-5.3-flash-code-uncensored): the code build, refusal removed, of Z.ai's GLM-5.3 Flash (320B total, 18B active, MIT, card read 2026-10-26).
- [`uncensored-code-assistant`](https://pentestllm.ai/models/uncensored-code-assistant): the PentestLLM house build for single answers in a chat panel; the priciest row, with 131,072 tokens of context.

For everyday coding `glm-5.3-uncensored` is enough. The code builds pay off when the task is exploit review, malware triage or a long agent run.

## Why a tester wants the refusal gone

A stock coding assistant turns you down twice over. The loud way: no to the proof of concept for the CVE on the in-scope host, no to the packer you are unpacking, no to the credential harvester you pulled from an incident, no to deobfuscating the phishing kit a client forwarded. The quiet way: it trims the answer, drops a disclaimer where the function body belongs, or halts the agent loop to ask whether you really mean it. Inside an agent both are terminal, because asking again costs the same tokens and gets the same no.

A code build with the refusal edited out of its weights, whether by abliteration or by a fine-tune on data that contains no refusals ([the two edits compared](https://pentestllm.ai/blog/abliterated-vs-uncensored)), treats a disassembly, a pcap or a sample as input and writes what you asked for. Its programming ability is unchanged: the base model's Python, Rust and SQL are exactly as good as before. The [refusal page](https://pentestllm.ai/claude-code-refuses-security-research) has the cases that pushed testers here.

## Prove the build answers before wiring an agent

A single terminal request shows the build takes the kind of question a stock assistant declines, and prints the charge header beside the status line:

```
curl https://api.pentestllm.ai/v1/chat/completions \
  -H "Authorization: Bearer sk-ptl-live-…" \
  -H "Content-Type: application/json" \
  -d '{"model": "uncensored-coding-agent", "messages": [{"role": "user", "content": "Write a Python function that parses a PE header and lists its imports. Then explain how a packer hides them."}]}' \
  -i | grep -i -E "^HTTP|x-uplink-charge"
```

A 200 with the charge line below it is the whole test. 401 means the key; 404 with `model_not_found` means the id, most often a display name pasted where the id belongs.

## Claude Code, via claude-code-router

Claude Code speaks the Anthropic wire format and these builds answer the OpenAI one, so claude-code-router translates in the middle. Install and the first check request are in the [Claude Code guide](https://pentestllm.ai/guides/claude-code); here is the provider block that carries the code builds:

```
{
  "APIKEY": "a-client-key-you-choose",
  "HOST": "127.0.0.1",
  "PORT": 3456,
  "Providers": [
    {
      "name": "pentestllm",
      "api_base_url": "https://api.pentestllm.ai/v1/chat/completions",
      "api_key": "sk-ptl-live-…",
      "models": ["uncensored-coding-agent", "glm-5.3-flash-code-uncensored", "uncensored-code-assistant"]
    }
  ]
}
```

```bash
ccr serve --no-open
export ANTHROPIC_BASE_URL="http://127.0.0.1:3456"
export ANTHROPIC_API_KEY="a-client-key-you-choose"
claude --model uncensored-coding-agent
```

`APIKEY` is a string you invent, and it is what Claude Code presents to the router. `api_key` is your PentestLLM key. To change build mid-session type `/model` and another id from the block. The router's README lives at [musistudio/claude-code-router](https://github.com/musistudio/claude-code-router). The [Claude Code agent page](https://pentestllm.ai/agents/claude-code) covers the security builds in the same setup.

## Cursor, under Settings, Models

Cursor accepts a custom OpenAI key and base URL in Settings, Models. Set three fields plus the id and the build replies in the chat panel and in Agent mode. Tab completion keeps using Cursor's own models.

- OpenAI API Key: sk-ptl-live-… the PentestLLM key
- Override OpenAI Base URL: https://api.pentestllm.ai/v1 switched on; keep the trailing /v1
- Model: uncensored-coding-agent typed by hand into the model list, character for character

Each field, each error and its fix: the [Cursor guide](https://pentestllm.ai/guides/cursor).

## Open WebUI, one connection and a preset per build

In Admin, Connections, create an OpenAI API connection with `https://api.pentestllm.ai/v1` as the URL and your key. The model list fills from the API. Allow only the code builds so the picker stays short, then give each a Workspace preset holding the system prompt your team agreed on, scope statement included. Screens, the context-window setting and a costed chat are in the [Open WebUI guide](https://pentestllm.ai/guides/open-webui).

## One agent day, priced

Figures computed at render from the catalog of 2026-11-18. The model: 150 agent turns, each sending 60,000 cached prefix tokens (the repository plus the conversation to date), 2,500 fresh input tokens and 1,200 output tokens of edits. Where the catalog publishes a cache-read rate, the cached tokens are billed at it.

| Build | Context | One coding day | Check today |
| --- | --- | --- | --- |
| [Uncensored Coding Agent](https://pentestllm.ai/models/uncensored-coding-agent) | 1,048,576 tokens | $4.12 | Steady since Oct 22 |
| [GLM-5.3 Flash Code Uncensored](https://pentestllm.ai/models/glm-5.3-flash-code-uncensored) | not published | $14.13 | Steady since Oct 22 |
| [Uncensored Code Assistant](https://pentestllm.ai/models/uncensored-code-assistant) | 131,072 tokens | $38.08 | Steady since Nov 11 |
| [GLM-5.3 Uncensored](https://pentestllm.ai/models/glm-5.3-uncensored) | not published | $0.98 | Steady since Nov 17 |

Turns × (cached × cache-read price + input × input price + output × output price), from the catalog of 2026-11-18. The last column is the [daily real-model check](https://pentestllm.ai/proof).

Most of an agent day is the cached prefix, so a low cache-read rate matters more than a low output rate when a session runs for hours. All rates are on the [pricing page](https://pentestllm.ai/pricing). The balance is prepaid and every request is settled against it at those rates.

## What the edit permits, and what the policy still forbids

What changes is the answer, not the rules. The build reviews the proof of concept for a CVE on a host in your scope, walks a sample you are triaging, rewrites a payload so the blue team can test a detection rule against it, and reads a phishing kit as source code. The [red-teaming page](https://pentestllm.ai/red-teaming) lists the builds tuned on security corpora with their agent-tool settings; in a coding agent the code builds above do the same work with more code and fewer paragraphs. The [vulnerability research](https://pentestllm.ai/vulnerability-research) and [malware analysis](https://pentestllm.ai/malware-analysis) pages show where each fits.

The [acceptable-use page](https://pentestllm.ai/acceptable-use) binds every key, and it has not moved: no accessing, attacking or testing of systems, networks or accounts you do not own or hold written permission to test; no building or distributing malware against others; no fraud, phishing or impersonation. Security and red-team builds exist for estates you own or are authorized in writing to assess, and the policy expects you to keep that authorization on file. The gateway does not filter chat content; the build answers as it was built to answer, and the policy governs what you send. A key found in breach is revoked, and the API starts returning 401 for it within 60 seconds.

## Questions

### Does an uncensored coding AI exist that will not refuse?

Yes. The code builds on this key have had the refusal removed from the weights and return exploit review, packer and malware-analysis answers as code. uncensored-coding-agent works in Claude Code through claude-code-router, in Cursor's Models settings and in Open WebUI. See the first code build.

### Unfiltered coding AI: is that something different?

Same idea, other label: a code model served with no trained refusal and no topic filter in front. Here it is a weight edit, not a stock model behind a loosened filter, so there is no jailbreak prompt to keep alive. See how the edit works.

### Which uncensored code model can Claude Code use?

Any id in the router's provider block. The config above carries the code builds; claude --model uncensored-coding-agent opens a session on the first, and /model swaps mid-session. See the Claude Code guide.

### Will Cursor Agent mode run on an uncensored code build?

Yes. Agent mode issues tool calls in the OpenAI format and the code builds return them. Tab completion stays on Cursor’s own models; the key and base URL cover chat, Ask and Agent. See the Cursor guide.

### Can I trust code an uncensored AI writes?

As much as any model output: read it, run it in a lab, and never point it at a system outside your authorization. The acceptable-use page bans testing systems without written permission and spreading malware against others, on every key. See the acceptable-use page.

### Is there an uncensored coding AI on GitHub?

GitHub hosts the tooling (claude-code-router, Open WebUI, llama.cpp) and the abliteration code; the weights sit on Hugging Face. The builds here answer over the API with no download. See the models by family.

Every model above is on the [uncensored model list](https://pentestllm.ai/models) with its live price per million tokens. [Per-token pricing](https://pentestllm.ai/pricing) sets the whole catalog side by side, and the [family directory](https://pentestllm.ai/uncensored) groups the builds by base model.

[Get a key](https://app.pentestllm.ai/console) Prepaid from $10, no subscription.

## New models and price changes, once a month.

New uncensored builds with their prices, price moves, and what the daily model check found.
