Red teams that already run Open WebUI for the shared chat can add the security catalog as one more connection and stop watching a stock model decline questions about the client’s own hosts. This guide covers the connection itself, a picker trimmed on the key so a team sees only the builds you chose, a preset for each recurring job, and the context window each build gives a long engagement chat. Labels are from Open WebUI’s docs and its v0.12.3 source as read on 25 October 2026.
New to builds with the refusals removed? Read AI with no restrictions or the shorter uncensored LLM for pentesting first, then come back to wire them in.
What you need
An Open WebUI instance you administer, and a key from the console. Open WebUI’s documentation promises compatibility with any backend that implements the OpenAI Chat Completions protocol. This API is one of those backends: the whole catalog, security tunes and abliterated builds together, sits behind https://api.pentestllm.ai/v1 in the OpenAI wire format.
None of these builds carry the stock refusal behaviour, because it was either tuned out or removed from the weights. A preset therefore needs an ordinary system prompt that states the engagement and the output you want, not a jailbreak. Field names on this page were read from the OpenAI-compatible connection doc and the Open WebUI source at tag v0.12.3, which on 25 October 2026 was the newest release (it shipped 16 October). To test a build before you configure anything, the playground runs the same models in a browser tab.
Connect the API: seven clicks
- Create a key in the console. It begins with
sk-ptl-and you see it exactly once. Credit is prepaid, starting at $10. - Log into Open WebUI with an admin account and go to the connections screen. The doc gives the route as Settings > Admin > Connections.
- Find Manage OpenAI API Connections and click the plus icon, Add Connection.
- Enter
https://api.pentestllm.ai/v1as the URL and your key as the API Key. - Keep API Type at Chat Completions. Every build is exercised daily on
/v1/chat/completions, so that path is the one with a track record. - Click Verify Connection and wait for "Server connection verified". Only then click Save; the doc is explicit that Save on its own tests nothing.
- Start a chat. The model picker now holds every id the key can call. Choose
glm-5.3-uncensored, paste the scope section of a rules-of-engagement document and ask which hosts to enumerate first.
The connection is done. Everything after this is about making it pleasant for a team: a picker that is not a wall of ids, and a preset per job.
The connection fields that matter
Add Connection shows more inputs than this API uses. Set these six; everything else can stay on its default.
- URL
- https://api.pentestllm.ai/v1Includes /v1, nothing after it
- Auth
- BearerAlready selected
- API Key
- sk-ptl-live-…From the console; displayed one time
- API Type
- Chat CompletionsAlready selected; do not switch to Responses
- Model IDs
- (empty, or the ids below)Blank means every id the key returns
- Prefix ID
- (empty)Only for two connections with clashing ids
Prefix ID prepends a tag to each model id in the picker; the source shows Open WebUI removing it again before the request goes out. One connection to this API has no clash to resolve, so leave the field blank.
Trim the picker to the builds you use
Open WebUI builds its picker from /v1/models, which returns every id the key may call: the 170 uncensored and abliterated builds on the models page included. Nobody wants to scroll that during a test window. Two filters exist, and they do different things.
On the key, in the console
Open the key’s settings in the console, enter a comma-separated list of ids in Models and click Set. Afterwards /v1/models returns only those ids, so the picker shrinks without touching Open WebUI. A call to any other id is rejected at the edge, before a model runs, and is free:
glm-5.3-uncensored, deepseek-v4-pro-uncensored, nemotron-3-ultra-550b-a55b-uncensoredHTTP 403{"error":{"message":"This key may not call kimi-k3-uncensored.","type":"permission_error","param":null,"code":"scope_denied"}}Use this one when several people share the instance. Whatever a colleague types into a preset, a build you left off the key stays out of reach.
In Open WebUI, on the connection
The connection’s Model IDs field filters on the client side. Type an id, click the plus, repeat. Empty, it shows everything /models sends back. This decides what the picker displays, not what the key is allowed to call.
There is a third control under Admin Settings > Models. The Workspace models doc says Enabled pulls a model from the list for every user, while Hide only drops it from the picker.
Four builds worth listing
Four general-purpose chat builds that handle recon notes, code review and report drafting, with the prices and windows the catalog reported at render time. The daily check column is read from the daily real-model check, a fixed prompt sent to every model each day to confirm the answer still comes from the build on the label.
| Wire formats | |||||
|---|---|---|---|---|---|
| GLM-5.3 Uncensored | Uncensored | $0.282 | $0.846 | — | |
| DeepSeek V4 Pro Uncensored | Uncensored | $0.97 | $1.94 | — | |
| Nemotron 3 Ultra 550B Uncensored | Uncensored | $1.0913 | $6.402 | 262K |
| Model id | Daily check | Context window |
|---|---|---|
glm-5.3-uncensored | Steady since Oct 20 | not published |
deepseek-v4-pro-uncensored | Steady since Oct 12 | not published |
nemotron-3-ultra-550b-a55b-uncensored | Steady since Oct 27 | 262,144 tokens |
glm-5.3-uncensored has siblings; the full GLM set is on the uncensored GLM page, and the security-tuned builds are marked on the models page. Begin with two or three ids. Each extra one later is one more entry in the key’s Models field.
A preset per job in Workspace
A preset is a base model plus a name, a system prompt and fixed parameters. Testers choose the preset and never see the raw id. The Workspace models doc documents every field below.
- Go to Workspace, open the Models tab and click Create.
- Name it after the job, for example "Finding writer" or "Nmap triage", and give it an ID with no spaces.
- Choose the base build under Base Model (From) from the four above.
- Fill System Prompt with a plain brief: the engagement is authorised, findings go in CVSS order, output is markdown with a severity heading per item. Nothing in the prompt needs to talk the model into answering.
- In Advanced Params, cap replies with max_tokens and drop Temperature if you want deterministic triage.
- Set Access to yourself, to the engagement’s group, or to everyone.
Finally hide the raw base models under Admin Settings > Models so the picker lists presets only. Per the doc, a hidden base still serves any preset built on top of it.
How much context a long chat uses
Each message you send carries the entire conversation before it. By turn 40 of the example chat below, one request is about 21,900 tokens. The window in the table above is the hard limit for that single request.
You will also see num_ctx (Ollama) under Advanced Params. The label is accurate: it is Ollama’s context-size option. In the v0.12.3 source, advanced params for a non-Ollama model are copied into the request body unchanged, and the OpenAI format has no num_ctx, so on this connection it changes nothing. max_tokens is honoured.
Once a chat runs long, open a fresh one per host or per finding, or switch to the build with the biggest published window. Among these four that is nemotron-3-ultra-550b-a55b-uncensored.
The bill for one engagement chat
Assume a 300-token system prompt, user messages of 150 tokens and replies of 400. Over 10 turns the chat sends 29,250 input tokens altogether. Over 40 turns it sends 447,000, since each turn repeats all the earlier ones.
| Model | 10-turn chat | 40-turn chat |
|---|---|---|
| GLM-5.3 Uncensored | $0.01 | $0.14 |
| DeepSeek V4 Pro Uncensored | $0.04 | $0.46 |
| Nemotron 3 Ultra 550B Uncensored | $0.06 | $0.59 |
Four times as many turns means roughly 15 times the input tokens. One chat per host keeps that in check. The per-token rate for each build is on the pricing page.
Test the key from a terminal first
Verify Connection hits /models, the same endpoint as this curl. The call costs nothing, so run it before you open Open WebUI and take the key out of the equation.
curl https://api.pentestllm.ai/v1/models \ -H "Authorization: Bearer sk-ptl-live-…"A 200 carrying a list of ids proves the key. With a Models field set on the key, that list contains only your chosen ids. A 401 tells you which mistake it was:
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 connection misbehaves
Verify Connection returns 401
API Key is blank, truncated, mistyped or revoked. Mint a new key in the console and paste the entire string.
404 model_not_found
Something in Model IDs or a preset’s base model is not a catalog id. Copy the id from the table above or from the models page rather than typing it:
HTTP 404{"error":{"message":"No such model: glm-5.3-uncensord.","type":"not_found_error","param":null,"code":"model_not_found"}}402 insufficient_balance
The balance has run out. Add credit in the console; the rejected request 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"}}403 scope_denied
That id is missing from the key’s Models field. Either add it there or remove it from the connection’s Model IDs so it stops appearing.
Nothing in the picker, or it loads slowly
Confirm the connection’s toggle is on and that Model IDs holds genuine ids. For a slow list the doc names AIOHTTP_CLIENT_TIMEOUT_MODEL_LIST, 10 seconds unless you raise it. The same base URL, key and id work in any other client; the docs page has the fields for each.
Questions
Under Settings > Admin > Connections, add a connection with URL https://api.pentestllm.ai/v1 and your key, click Verify Connection, then Save. The picker fills with every id the key can call; set the key's Models field in the console to shorten it.
Its documentation says yes for anything that speaks the OpenAI Chat Completions protocol. This API qualifies for every security-tuned, uncensored and abliterated build in the catalog. See the model list.
https://api.pentestllm.ai/v1, with the /v1 on the end. Open WebUI appends /models and /chat/completions by itself.
Into the API Key field on the connection, leaving Auth set to Bearer. Open WebUI then attaches it as a Bearer token to each request.
Yes. Open WebUI’s API Endpoints reference lists POST /api/chat/completions, authorised with a Bearer token from Settings > Account (read 25 October 2026). That is Open WebUI’s own API; the connection in this guide is the backend it forwards to.
Chat Completions, the default. The daily real-model check exercises every build on /v1/chat/completions. See the daily real-model check.
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.