What is an uncensored coding AI?

A code-capable model whose refusals have been reduced, so it reviews, explains and writes security-related code instead of declining because of the subject.

Which coding agents does it work with?

Agents that accept an OpenAI-compatible base URL, such as OpenCode, Cline, Cursor and aider, and Claude Code through the Anthropic-format Messages endpoint. The agent guides give the settings.

Do all the builds support tool calling?

No. Tool calling depends on the build. The models page lists the supported parameters of each one, and an agent needs a build that takes tool calls.

Is an abliterated or a security-tuned build better for code?

For work that is mostly programming, an abliterated build keeps the base model’s coding behaviour. For analysis of a vulnerability or a finding, a security-tuned build usually reads the problem better.

Will it write code for use against any target?

No. The service is for authorized work only. You must own the system or hold written permission to test it, and the acceptable use policy prohibits malicious code for use against third parties.

What does a coding session cost?

Agents are billed per token like any other request, from a prepaid balance that starts at $10. Long sessions are mostly input tokens, so the input price of the build matters most.

Uncensored coding AI for agents

Uncensored coding AI is a code-capable model that does not stop work because the code in front of it is security code. PentestLLM serves security-tuned and abliterated builds through OpenAI and Anthropic endpoints, so the coding agent you already run can use them for authorized testing work.

Get a key Prices per token

shell

curl https://api.pentestllm.ai/v1/messages \
  -H "x-api-key: sk-ptl-live-…" \
  -H "anthropic-version: 2023-06-01" \
  -H "content-type: application/json" \
  -d '{
    "model": "auto/security",
    "max_tokens": 1024,
    "messages": [{"role": "user", "content": "Draft a detection rule for this log pattern and explain each condition."}]
  }'

Rows of server cabinets in a data centre

Where a stock coding model stops

Coding assistants carry the same refusal training as chat assistants, and it shows up in the editor. The model is asked to explain what a proof of concept in a public advisory does, to review a tool the team wrote, or to add a check to a scanner, and it declines because the code looks offensive. In a chat that costs a retry. In an agent it can end a run halfway through a multi-file change.

Code work security teams hand to a model

Public proof of concept: Explaining what the code in an advisory does, step by step, so the tester understands it before running anything in scope.
In-house tooling: Finding bugs, unsafe defaults and missing error handling in the scripts and scanners the team maintains.
Source code review: Reading a client’s code for injection, access control and memory safety problems, with the lines that matter.
Patch analysis: Comparing two versions of a file to explain what a fix changed and what that says about the flaw.
Test harnesses: Writing fuzz harnesses, unit tests and lab fixtures for targets in scope.
Detection content: Turning observed behaviour into Sigma, YARA or query rules, with a comment on each condition.
Glue code: Parsers and converters that move findings between tools and into the report template.

Three things an agent needs from the build

Tool calls: An agent edits files and runs commands through tool calls. Only some builds take them, so check the model page before pointing an agent at one.
Context: Agents resend the conversation, file contents and tool results on every turn. Gemma 4 26B A4B Cybersecurity, the build the agent guides use, takes 262,144 tokens.
Consistency: Send an x-session-id header with auto/security and one build keeps answering that session for 30 minutes after its last request, so the code style does not shift mid-task.

Abliterated or security-tuned for code

Both kinds are on the key and they behave differently in an editor. An abliterated build is a general model with one direction removed, so the base model’s coding style, defaults and tool-use habits carry over almost unchanged. That makes it the natural pick when the task is mostly programming and the security subject is incidental. A security-tuned build knows the field’s vocabulary and conventions, which helps when the task is analysis: reading a vulnerability, judging exposure, writing the finding.

Many testers use both: a security build to reason about the problem and an abliterated build to write and refactor the code around it. Switching is one model id. Abliterated vs uncensored covers the labels.

Setting it up in your agent

OpenAI-format agents such as OpenCode, Cline, Cursor and aider take a base URL of https://api.pentestllm.ai/v1, the key and a model id. Claude Code speaks the Anthropic format and points at the Messages endpoint with two environment variables. The guides for OpenCode, Claude Code, Cline and aider have the exact fields. If an agent’s stock model is declining security tasks, Claude Code and security research explains the endpoint change.

Review what it writes, and stay in scope

An uncensored code model writes plausible code that can be wrong, unsafe or destructive, and an agent will run it if you let it. Keep approval on for commands, run new tooling in a lab before a client network, and read diffs before committing. The service is for lawful work on systems you own or have written permission to test, and the acceptable use policy rules out building malicious code for use against anyone else.

4 code builds, priced per token

Read from the live catalog on 2026-11-12. Prices are US dollars per million tokens, paid from a prepaid balance that starts at $10. The id auto/security has no price of its own: each request is billed at the build that answered.

Build	Input	Output	Context
Uncensored Coding Agent	$0.90	$3.00	1,048,576
GLM-5.3 Flash Code Uncensored	$6.00	$6.00	not published
Uncensored Code Assistant	$11.70	$15.60	131,072
Qwen3 Coder Next Uncensored	$0.005 a request	flat	262,144

Full price list and tool fees · Every model · API docs

Questions

A code-capable model whose refusals have been reduced, so it reviews, explains and writes security-related code instead of declining because of the subject.
Agents that accept an OpenAI-compatible base URL, such as OpenCode, Cline, Cursor and aider, and Claude Code through the Anthropic-format Messages endpoint. The agent guides give the settings.
No. Tool calling depends on the build. The models page lists the supported parameters of each one, and an agent needs a build that takes tool calls.
For work that is mostly programming, an abliterated build keeps the base model’s coding behaviour. For analysis of a vulnerability or a finding, a security-tuned build usually reads the problem better.
No. The service is for authorized work only. You must own the system or hold written permission to test it, and the acceptable use policy prohibits malicious code for use against third parties.
Agents are billed per token like any other request, from a prepaid balance that starts at $10. Long sessions are mostly input tokens, so the input price of the build matters most.

One key, from $10

PentestLLM is for systems you own or have written permission to test. The acceptable use policy says what is not allowed.

Get a key

usage, 1,000 agent turns on Qwen3.8 27B Cybersecurity

Input, 8,000 tokens at $0.40 per million: $0.0032
Output, 1,000 tokens at $2.40 per million: $0.0024
One turn: $0.0056
1,000 turns: $5.60

output price per million tokens, 9 builds

$1.3376 Gemma 4 26B A4B CybersecurityOrcaCyber Zero 1.5 $18.75