A security engineer who automates triage in n8n, scanner output in, ranked findings out, hits the same wall as everyone else: the stock model behind the AI Agent declines once the payload mentions an exploit, even on hosts the client signed for. Because n8n’s OpenAI credential accepts a custom Base URL, any OpenAI-compatible API can sit behind it, and this guide puts a security-tuned or abliterated build there: the credential, the Chat Model sub-node hung off an AI Agent, the model id typed by hand, and the errors that come up along the way. Field names are from n8n’s source as read on 25 October 2026.
If builds with the refusals removed are unfamiliar, the short explainer is abliterated vs uncensored; come back here afterwards to wire one in.
Two meanings of the phrase, one covered here
Search for "n8n OpenAI-compatible API" and you get two unrelated problems mixed together. One is pulling an external model into a workflow. The other is publishing a workflow so other software can call it as if it were OpenAI. Only the first is on this page: a security-tuned or abliterated build as the model behind n8n’s AI nodes, so a scheduled triage workflow does not stall on a refusal.
Field labels were taken from the n8n source on GitHub on 25 October 2026, when n8n@2.44.6 was the current release. The docs for these nodes did not yet mention the Base URL field at that point; the source did.
What you need
An n8n instance, a key from the console, and n8n’s OpenAI credential, which has four inputs. Set them as follows:
- API Key
- sk-ptl-live-…From the console; displayed one time
- Organization ID (optional)
- (empty)Ignored by this API
- Base URL
- https://api.pentestllm.ai/v1Overrides the OpenAI default; keep the /v1
- Add Custom Header
- offLeave it off
On save, n8n probes the credential with a GET to /models beneath the Base URL. That request is free here, and it is exactly the curl further down. From node version 1.1 onwards the OpenAI Chat Model reads its Base URL from the credential; there is no option on the node for it.
Shortest path to a first run
- Make a key in the console. It starts
sk-ptl-. Credit is prepaid from $10. - Add a trigger. A chat trigger is handy for testing; later it becomes the webhook that receives scanner output. Follow it with an AI Agent node.
- Attach an OpenAI Chat Model sub-node to the agent’s Chat Model connector.
- In the sub-node’s credential field, create a fresh OpenAI credential using the four values from the table, then save it. The test should go green.
- Under Model, change the mode from From List to ID, then paste in an id such as
glm-5.3-flash-uncensored. - Switch Use Responses API off.
- Optionally open Options to cap Maximum Number of Tokens or lower Sampling Temperature.
- On the agent, write the brief into the System Message option, for example: the input is nuclei output from an authorised engagement; group findings by host, rank by severity, draft one paragraph per finding. Run it.
The two settings most people skip are ID mode and the Responses switch. Each gets its own section next. To check a build’s output before building the workflow, the playground runs the same model with no install.
Why to type the id rather than pick it
Against the stock OpenAI URL, n8n narrows the From List dropdown to OpenAI’s chat models. Against any other Base URL, the node’s model search shows everything /models returns, which on this API is every id the key may call, the 170 uncensored and abliterated builds among them. You can search that list, but ID mode is faster and cannot be misclicked.
It is also the fix a community thread on precisely this setup arrived at. If you would rather shorten the list, open the key in the console and put the ids you use in Models. /models then returns those alone, and the key is barred from calling anything else.
Switch off Use Responses API
Node version 1.3 of the OpenAI Chat Model ships with Use Responses API enabled. Left on, n8n speaks OpenAI’s Responses format rather than Chat Completions.
Disable it. The daily check exercises every build at /v1/chat/completions, which makes Chat Completions the format with a track record. The Built-in Tools beneath that switch, web search, file search and code interpreter, execute on OpenAI’s infrastructure and play no part here.
Agent, tools, or a Basic LLM Chain
In the current AI Agent (node version 3.1 in the source) the only connector that must be filled is Chat Model. Memory and Tool are optional, so an agent without tools behaves as a single chat step. The AI Agent docs page still asks for at least one tool. Should your n8n version insist, the Basic LLM Chain accepts the identical OpenAI Chat Model sub-node and needs no tools.
Tools only pay off when the build handles tool calls reliably. The node shows its own notice once the Base URL is overridden, to the effect that tool calling and JSON output are not guaranteed on a model from outside OpenAI. Run one tool call, say a lookup against your own CVE database, on the chosen build before a workflow relies on it.
One message that should never appear with this setup: "Tools Agent requires Chat Model which supports Tools calling". The source raises it only when the attached sub-node has no tool binding whatsoever, and the OpenAI Chat Model always carries one.
Three builds for unattended runs
A workflow fires often and nobody is watching, so the price of one run matters more than it does in a chat window. Three candidates, with the catalog figures at render time and today’s verdict from the daily real-model check, which sends one fixed prompt to each model every day:
| Wire formats | |||||
|---|---|---|---|---|---|
| GLM-5.3 Flash Uncensored | Uncensored | $0.1053 | $0.3685 | — | |
| GLM-5.3 Uncensored | Uncensored | $0.282 | $0.846 | — | |
| DeepSeek V4.1 Flash Abliterated | Abliterated | $0.8488 | $1.6975 | — |
| Model id | Daily check | Context window |
|---|---|---|
glm-5.3-flash-uncensored | Steady since Oct 12 | not published |
glm-5.3-uncensored | Steady since Oct 12 | not published |
deepseek-v4.1-flash-abliterated | Steady since Oct 19 | not published |
Put the high-volume steps, such as classifying every alert from a bug bounty program, on the flash build and promote a step to the full build only when its output disappoints. Swapping is one field, the id in the sub-node. The rest of the GLM line is on the GLM family page, and auto/security on the auto router picks a security build per request by difficulty.
The bill for 1,000 runs
Take 1,000 runs, each sending roughly 1,500 tokens of system message plus trigger payload and receiving a 400-token answer.
| Model | 1,000 runs | 30,000 runs |
|---|---|---|
| GLM-5.3 Flash Uncensored | $0.31 | $9.16 |
| GLM-5.3 Uncensored | $0.76 | $22.84 |
| DeepSeek V4.1 Flash Abliterated | $1.95 | $58.56 |
An agent that uses tools issues several model requests per run, and each one carries the conversation so far. Scale the figure by the number of steps your agent typically takes. Per-token rates for every build are on the pricing page.
Test the credential without n8n
Below is the request n8n’s credential test performs. It is free: run it in a shell, or build an HTTP Request node with that URL and that header.
curl https://api.pentestllm.ai/v1/models \ -H "Authorization: Bearer sk-ptl-live-…"A 200 carrying a list of ids confirms the key and the Base URL together. A 401 tells you how the key is wrong:
HTTP 401{"error":{"message":"No API key was sent.","type":"authentication_error","param":null,"code":"invalid_api_key"}}HTTP 401{"error":{"message":"The API key is not valid.","type":"authentication_error","param":null,"code":"invalid_api_key"}}When the credential or the node fails
The credential test goes red
Start with the Base URL: https://api.pentestllm.ai/v1, /v1 and nothing more. Then the key: it must come from this console, not another vendor’s dashboard. That confusion was the opening reply in the community thread linked above.
The model dropdown shows No results
The dropdown is populated from /models under the Base URL, so a URL missing /v1 empties it. Switch to ID mode and the dropdown is irrelevant.
404 model_not_found
The id in Model does not exist in the catalog. Paste it from the curl output or from the models page.
HTTP 404{"error":{"message":"No such model: glm-5.3-flash-uncensord.","type":"not_found_error","param":null,"code":"model_not_found"}}402 insufficient_balance
The balance is empty. Add credit in the console; the rejected run was not billed.
HTTP 402{"error":{"message":"The prepaid balance does not cover this request. Top up at … to continue.","type":"insufficient_quota","param":null,"code":"insufficient_balance"}}"Bad request - please check your parameters"
n8n falls back to this message when it cannot parse the error. Its page on common issues recommends repeating the call in an HTTP Request node to read the whole body. The same Base URL, key and id serve every other client; the docs page has their fields.
Questions
The OpenAI credential, under API Key and Organization ID. The default is https://api.openai.com/v1; swap it for https://api.pentestllm.ai/v1.
Yes, by way of the credential’s Base URL. Once that points elsewhere, the OpenAI Chat Model lists whatever ids the endpoint returns and takes any id typed in ID mode. See the model list.
Anything that speaks the OpenAI format can sit behind the Chat Model sub-node. This API serves security-tuned, abliterated and uncensored builds, charged per token against a prepaid balance that does not expire. See per-token prices.
Not in the current node: n8n’s source requires only the Chat Model connector. If your version demands a tool, the Basic LLM Chain takes the same sub-node without one.
A community project for the reverse direction: it exposes n8n workflows to other software as if they were an OpenAI endpoint. This guide covers calling a model from inside n8n.
Every model above is on the uncensored model list with its live price per million tokens. Per-token pricing sets the whole catalog side by side, and the family directory groups the builds by base model.