When a no filter AI chat app says no filter, it means the filter it runs. The one trained into the model is still there. The way around it is an abliterated model: refusal removed at the weights, no app layer to argue with. You can try one in the browser right now, or connect it to the chat client you already use with a base URL, a key and a model id. This page is the client-side companion to AI with no restrictions.
Try it first: no filter AI chat with nothing to install
The browser chat runs the same models this page connects to. You get three free turns a day with no account: no email, no phone, no ID. Settings holds the system prompt and the temperature. Files, web search and characters are on the same screen. The conversation stays in your browser; the server keeps no copy of what you typed.
Prefer a phone? The Telegram bot answers three messages a day for free. Everything below is for the moment you want your own client, your own character cards and a model you picked by name.
The three settings every client asks for
A chat client with a custom OpenAI-compatible option needs a base URL, a key and a model id. Nothing else. Make the key in the console; it is displayed one time, so copy it then. The id in the list is an abliterated build, which means the refusal was cut out of the weights and no system prompt has to fight it.
- Base URL
- https://api.unrestrictable.ai/v1
- Key
- sk-unr-live-…Made in the console, displayed one time
- Model
- glm-5.3-flash-abliteratedGLM-5.3 Flash Abliterated, context not in the catalog
Test the key in a terminal before you open a client. One call, one reply, one line on the balance.
curl https://api.unrestrictable.ai/v1/chat/completions \ -H "Authorization: Bearer sk-unr-live-…" \ -H "Content-Type: application/json" \ -d '{ "model": "glm-5.3-flash-abliterated", "messages": [ {"role": "system", "content": "You are the narrator of a heist story set in 1970s Marseille."}, {"role": "user", "content": "Open the scene at the harbour, two paragraphs."} ], "stream": false }'The reply uses the standard chat completions shape and ends with a usage block. Any other id on the Models page drops into the same field, and its per-token rate is printed beside it. For the difference between these builds and a jailbreak prompt, read AI with no restrictions.
Each build on the Models page carries a label. Abliterated: the refusal direction was projected out of the weights; heretic and derestricted builds count here too. Uncensored: a fine-tune or merge with no reflex to decline. Roleplay: trained on long stories and character writing, and it holds a persona through a whole session. Security: takes red-team and code-analysis prompts a stock model rejects. In a chat client, begin with an Abliterated or Roleplay build; the rest can wait for a second tab.
Where the refusal in a no-filter app comes from
An answer can be blocked at two points. Point one is the layer a service wraps around the model: a classifier reading your message, another reading the reply, a word list. An app can turn that layer off and put the fact in its tagline. Point two lives in the weights. Alignment training taught the model a refusal direction, and that direction fires by itself, no matter what the app does. So a chat that advertises no filter can still meet a story prompt with a sermon.
Abliteration deals with point two. It locates the activation direction that splits refused prompts from answered ones, then removes that direction from the weights. Maxime Labonne’s Hugging Face write-up shows every step on an 8B model. What remains is the same model, same knowledge, same prose, minus the trained habit of declining. Uncensored builds get there by a different road: they are fine-tuned on answers rather than refusals, or merged with a model that was.
A hosted app also chooses the model for you, and the one behind a persona is seldom named. Set up the client yourself and that reverses: you type the id, and every morning the daily real-model check puts one fixed prompt to each id and writes down whether the reply moved. On November 16, 2026 it reported 146 of the 168 listed models stable.
The check has no tricks in it. The same prompt goes to each model id at the same hour daily. The reply is set against the replies on file for that model. Match: stable, dated from when it last changed. Drift: changed, with the date. Silence: unchecked. The page says nothing about who hosts the model, only whether the id in your client still acts like the model it is named for. An app that hides its model cannot offer that.
SillyTavern setup for no filter AI chat
In SillyTavern the connection goes through the Custom (OpenAI-compatible) source. Presets and context size are in the full SillyTavern guide; the connection steps follow.
- Open the API Connections panel.
- Set API to Chat Completion.
- Set Chat Completion Source to Custom (OpenAI-compatible).
- Endpoint URL:
https://api.unrestrictable.ai/v1. - API key: paste in the key.
- Press Connect. Choose
glm-5.3-flash-abliteratedin the model list. If the list stays empty, type the id into the model field. Send one test message.
Keep the preset’s context size at or under the model’s window. The catalog lists no window for glm-5.3-flash-abliterated; leave the preset’s context size where it is.
Janitor AI setup for no filter AI chat
Janitor AI talks to an outside model through a proxy configuration. Two things differ from other clients: it takes the full chat completions path rather than the base URL, and the model id is typed, not picked. The field names are taken from the proxy quickstart Janitor publishes, read October 25, 2026. Errors and context size are in the Janitor AI proxy guide.
- Open any chat and click the gear at the top right. Choose API Settings.
- Press Add Configuration.
- Config Name: whatever you like. Model Name:
glm-5.3-flash-abliterated, spelled as the Models page spells it. - Proxy URL:
https://api.unrestrictable.ai/v1/chat/completions. - API Key: the key, with no spaces around it. Custom Prompt stays blank.
- Press Add, then Save Settings. Reload Janitor and send a short test line.
A 401 points to a key pasted with a stray space or a lost character. An error that repeats the model name back to you points to a typo in the id. Nothing is charged for a request that fails before the model replies.
Open WebUI and phone clients
Save the connection in Open WebUI and the whole catalog appears in its model picker.
- Settings, Admin, Connections.
- Find Manage OpenAI API Connections and press the plus sign beside it for a new connection.
- URL:
https://api.unrestrictable.ai/v1. API Key: the key. - Save. Start a new chat, choose
glm-5.3-flash-abliteratedin the picker and send something.
On iOS or Android, pick any client with a custom OpenAI-compatible option and give it the same trio: base URL, key, model id. The wire format is OpenAI’s public chat completions specification, and a client written against that spec works with no plugin. Scripting it instead? The docs page has Python and Node snippets.
The labels vary by client. Whether the field says Base URL, API host, endpoint or custom provider, it wants https://api.unrestrictable.ai/v1: the /v1 stays, nothing follows it. A client that asks for a full path, as Janitor AI does, wants https://api.unrestrictable.ai/v1/chat/completions. Leave any organization or project id field empty. If there is a button that fetches the model list, press it. The ids that come back match the Models page, and a filled list confirms the URL and the key before you spend a message.
The bill for 100 and 1,000 messages
No subscription is required. You top up the balance from $10.00 by card, WeChat Pay or crypto, with no ID check, and each request takes its tokens out of it. Credit keeps for twelve months. Three builds, pulled from the live catalog when this page renders:
| GLM-5.3 Flash Abliterated | Abliterated | $0.4656 | $1.94 |
| Cydonia 24B v4.3 | Roleplay | $0.484 | $0.60 |
| Llama 3.3 70B Abliterated | Abliterated | $2.80 | $2.80 |
The example that follows assumes 1,500 input tokens a message (the character card plus the recent history your client sends again each turn) and 250 output tokens in return, roughly three or four paragraphs. A long chat with a large context setting costs more per message; a short assistant reply costs less. The numbers come from the rates in the table above and are worked out as the page renders.
| Build | One message | 100 | 1,000 |
|---|---|---|---|
| GLM-5.3 Flash Abliterated | $0.0012 | $0.12 | $1.18 |
| Cydonia 24B v4.3 | $0.0009 | $0.09 | $0.88 |
| Llama 3.3 70B Abliterated | $0.0049 | $0.49 | $4.90 |
After a few turns, input tokens are the bill. The client sends the entire visible history every turn, so at full context a chat ships thousands of tokens to receive a few hundred. Two settings lower the per-message figure: a smaller context size, set in the client, and a summary extension, which compresses old turns into one paragraph. Both trim the input and leave the output alone, and output is where the builds in the table differ most.
Bonus credit on larger top-ups and the per-million rate for every model are on the pricing page. A spend limit per key is set in the console, so a client you share cannot go past the figure you set.
When something looks wrong
- The reply stays in character. A roleplay tune keeps the persona on its card, and that can look like an evasion. Switch the id to a plain abliterated build such as
glm-5.3-flash-abliterated, get the straight answer, switch back. - The reply stops mid-sentence. Your client caps the response length below what the model wanted to write. Raise the cap there; the API’s
max_tokenstakes whatever the client sends. - Two keys at once. If the client sets
Authorizationand alsox-api-key, and the two values disagree, the answer is a 400. Fill one field only. - The key in the URL. Rejected with a 400. The key goes in a header only, so it never reaches a server log or your browser history.
- Where the prompts go. Request bodies are not stored. The one exception is a key whose owner turned on the seven-day replay window in the console. This morning’s chat is not sitting on a disk this afternoon.
- The id is gone. The catalog moves. The Models page shows what is live now, and the check page shows which ids replied this morning; choose from those.
- A wait before the first word. When
streamis on, the first token still has to wait for the model to read the whole history, so more context means more waiting. A smaller context setting cuts it. - One key, two people. A key accepts requests from any number of clients at once, all against one balance. For the second person, create a second key in the console; each then carries its own spend limit and can be revoked separately.
If you never want to see a base URL
The hosted apps in the search results have their place. Venice puts chat, image and video generation on one screen and opens its home page with privacy (read October 25, 2026). Several hosted character-chat apps also held first-page spots for this query on Google when we looked (October 24, 2026). If you want to be typing within thirty seconds, use a hosted app, and ours asks for no account to start.
Wiring up your own client costs one key and three fields. What you get back: the model chosen by name, your existing cards and presets untouched, the per-token rate printed beside every id, and a daily check that says the id you chose answered this morning.
Questions
Yes. The browser chat gives three free turns a day with no account, email or phone number. For your own client you need a key, which takes an account and a prepaid balance from $10.00; you then pick the exact model and pay per token, with no subscription. See free AI chat, no sign-up.
Yes: an abliterated open-weight model served through the same chat completions interface ChatGPT clients already speak. With the refusal gone from the weights, the client, the prompt format and streaming all work as before, and no jailbreak prompt is needed. See AI with no restrictions.
Chat requests are not screened by topic. What is never allowed is written in the acceptable use policy. Image and video prompts that involve minors are refused. See the acceptable use policy.
Whichever runs the model you would have chosen yourself. Sort the Models page by price or context window, pick an id, confirm it on the daily real-model check, then connect it in SillyTavern, in Janitor AI, in Open WebUI, or simply open it in the browser chat. See every uncensored model, priced.
Yes. The browser chat works on a phone as it is. Any Android client with a custom OpenAI-compatible option takes the base URL, the key and a model id and sends the same request a desktop client sends.
Per token, no monthly fee. The table on this page works out 100 and 1,000 messages for three live builds, using the catalog rates at render time; the pricing page covers every model. See per-token prices.
No. A jailbreak is a prompt that argues a filtered model past its training and breaks when the model is updated. An abliterated model has the refusal cut from the weights, so there is no trick to maintain. See abliterated vs uncensored.
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.