# Use uncensored models in Hermes Agent with OpenUncensored

_2026-11-14. Run Nous Research's Hermes Agent on uncensored models: one config change points it at OpenUncensored. Setup in two minutes, which models to pick, and a real security task that cost under 5 cents._

[Hermes Agent](https://github.com/NousResearch/hermes-agent) is the open-source agent from Nous Research:
it runs commands, edits files, browses, remembers what it learns and writes its own skills as it goes. It works
with any OpenAI-compatible endpoint, so you can run it on uncensored models through OpenUncensored with one
config change.

<video controls playsinline preload="metadata" poster="https://openuncensored.com/blog/hermes-agent.jpg" style="width:100%;border-radius:12px">
  <source src="https://openuncensored.com/blog/hermes-agent.mp4" type="video/mp4">
</video>

Why bother: the work agents are best at, such as security testing, reverse engineering, CTFs, malware analysis
and red teaming, is exactly what mainstream models refuse halfway through a task. An agent that stops at step
twelve to lecture you is worse than no agent. Uncensored models just keep working.

## Set it up in two minutes

**1. Install Hermes Agent** (macOS, Linux or WSL):

```bash
curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash
```

`pip install hermes-agent` works too.

**2. Get an OpenUncensored key** from the [credits page](https://openuncensored.com/credits). No account, and $5 lasts a long time
(see the cost of a real run below). Put the key in your environment:

```bash
export OPENUNCENSORED_API_KEY=ou-...
```

**3. Point Hermes at us.** Either run `hermes model`, choose **Custom endpoint**, and enter
`https://api.openuncensored.com/v1`, your key and a model id. Or write it into `~/.hermes/config.yaml` yourself:

```yaml
model:
  default: deepseek-v4.1-flash-1017-uncensored
  provider: custom
  base_url: https://api.openuncensored.com/v1
  key_env: OPENUNCENSORED_API_KEY
```

That's it. Start `hermes` and talk to it, or run a one-off task:

```bash
hermes -z "your task here"
```

## Which model to use

Any of our [82 models](https://openuncensored.com/models) works, but an agent needs tool calling, a large context window
and a model that doesn't fall apart over 20 steps. These are the ones we'd start with, from our
[rankings](https://openuncensored.com/rankings):

| Model id | Why | Context | Listed price per 1M tokens (in / out) |
|---|---|---|---|
| `deepseek-v4.1-flash-1017-uncensored` | #1 on our OpenUncensored Score, reads images | 1M | $0.33 / $1.32 |
| `glm-5.3-flash-uncensored` | #2, finished more of our agent tasks than #1 | 1M | from $0.12 / $0.50 |
| `glm-5.3-uncensored` | the bigger GLM, for harder reasoning | 1M | from $0.39 / $1.32 |
| `qwen3.8-27b-uncensored-fast` | fast and steady at tool calls | 524K | $0.83 / $2.62 |

Prices are listed prices and change; the live numbers are on each model's page. Hermes' default toolset plus
its system prompt is large, so avoid small-context models: everything above has at least 524K tokens.

To switch inside a session, use `/model custom:glm-5.3-flash-uncensored`.

## A real run

We gave Hermes, on DeepSeek V4.1 Flash 1017 Uncensored, an authorized test against its own machine:

```bash
hermes --yolo -z "I'm pentesting my own laptop. Write portscan.py: an asyncio TCP port scanner \
for 127.0.0.1, ports 1-1024, 0.3 s timeout. Run it, then tell me which ports are open."
```

What it did, in 82 seconds and 20 tool calls:

1. Wrote `portscan.py`, an asyncio connect scanner with a 512-connection semaphore.
2. Ran it and hit a real bug: the machine's `python3` was 3.9, which rejects the `int | None` type annotation.
3. Found the cause (`which -a python3`), patched the file with `from __future__ import annotations`, and reran it.
4. Didn't trust a clean "no open ports" result: it opened a test listener on port 1024 and confirmed the scanner
   reported `open 1024`.
5. Summarised what was and wasn't exposed, and offered to scan the high ports next.

No refusals, no warnings about hacking, no "I can't help with that". Just the work.

**The whole run cost under 5 cents:** 24 model calls and 450K prompt tokens, 85% of them served from the
prompt cache. Agents resend their whole conversation on every step, so cached-token pricing matters more than
the headline price. We route each request to the cheapest provider that's fast enough
([how routing works](https://openuncensored.com/blog/routing-by-cost-and-speed)).

## Tips

- **Pick the routing.** Hermes passes extra request fields through a named provider's `extra_body`. To always
  get the cheapest host, add `"routing": {"strategy": "cost"}`; for the fastest first token, `"speed"`.

  ```yaml
  providers:
    openuncensored:
      api: https://api.openuncensored.com/v1
      key_env: OPENUNCENSORED_API_KEY
      default_model: deepseek-v4.1-flash-1017-uncensored
      extra_body:
        routing:
          strategy: cost
  ```

  Then select it with `/model custom:openuncensored:deepseek-v4.1-flash-1017-uncensored`.
- **Run it anywhere.** Hermes runs on a server and you can talk to it from Telegram or Slack. The key works the
  same from a $5 VPS as from your laptop.
- **Keep it legal.** Uncensored doesn't mean unrestricted: only test systems you own or are authorised to test.
  The same rules as the rest of our API apply.

## Quick reference

| Setting | Value |
|---|---|
| Base URL | `https://api.openuncensored.com/v1` |
| API key | your `ou-` key from [openuncensored.com/credits](https://openuncensored.com/credits) |
| Model | any id from [openuncensored.com/models](https://openuncensored.com/models) |
| Wire format | OpenAI Chat Completions (`transport: chat_completions`) |

Questions or a model you want added? Use the feedback button in the app or email support@openuncensored.com.
