Run Hermes Agent privately on your Mac with the Venice API

What this build is and installing Hermes

The premise is simple: Hermes Agent recently added computer use, so it can now control your machine. The question this guide answers is which model drives it and how to keep the whole setup private. The plan is a fresh install on a Mac, connected to a Venice model that does not store your data, used to organize a desktop, create reminders, and build a small chat widget.

Installation is a single command. Copy the install command from the Venice docs, open the Terminal app on your machine, and paste it. The installer script pulls down everything Hermes needs, including Python, Git, and Node.js, then automatically launches the Hermes setup. You do not run the configuration step yourself, the script does it for you. Choose the quick setup when prompted.

This is brand new software, so expect some finicky behavior depending on your machine. Treat it as a skill worth practicing now so you are ahead of the curve as agents become normal.

Connecting Venice as a custom provider

During setup, add Venice as a provider. If Venice is not yet in the provider list, choose the custom endpoint option and enter the API base URL:

https://api.venice.ai/api/v1

You can copy this exact URL from the API section at docs.venice.ai. Submit it, then generate an API key from the API section of the Venice interface: click generate new API key, name it something like "Hermes Agent", create it, copy it, and paste it into the terminal.

If you are new to Venice, the DIEM token is worth knowing about. Once you stake it, you receive a recurring $1 of Venice API credit every single day, which adds up when you are running an agent that makes frequent calls.

Reading the model table and picking one

After the key is accepted, Hermes pulls the available Venice models. You can see the same list under Models, then Text in the docs, with columns for whether each model is private, anonymized, or intent encrypted, plus its context length and supported tools (function calling, reasoning, vision, code optimization).

The goal is a model that does all of these things while staying cheap. The video compares options: Kimi K2.6 handles everything with a 256,000 token context at roughly $0.85 input and $4.66 output per million tokens, far below a frontier model like GPT 5.5 at around $6.25 and $37.50, with comparable results for this kind of work. Other strong all-rounders mentioned are DeepSeek v4 and Qwen 3.6.

The guide settles on Qwen 3.6 27B because it covers every capability and is even cheaper than Kimi. In the list it appears as option 13, so type 13, let Hermes auto-detect the token length, and name the provider "Venice". The takeaway: do not assume you need Opus or GPT for everything, test the open-source models first.

Keeping it local and private

Next, Hermes asks for a terminal backend, meaning where it runs commands. You can run locally on your machine, spin up an isolated Docker container, run on a cloud service, or even target a remote VPS or Raspberry Pi. For a private personal setup, choose to run locally.

Skip the messaging platform step. Connecting Telegram, Discord, or WhatsApp inserts a middleman that stores your messages and is not intent encrypted. Running Hermes directly on your computer means the only network hop is a direct connection through Venice to the model with zero data retention. That is the whole point of this configuration.

Letting Hermes optimize itself and defining its job

The default Hermes install loads many tools and skills, which inflates your context window even when unused. On first launch you will see it has already burned through about 16,000 tokens just loading everything. Fix this immediately with a prompt like:

Help me clean up the tools and skills we won't use.

Then define the role. The video uses framing like: "I want you to be my computer manager. You're not doing my work for me, but you're helping me stay organized: my files, my desktop, etc. Think of me as your personal digital janitor and organizational assistant." Hermes suggests which tool sets to keep or disable. Keep skills (so it can create its own skills for new tasks), keep code execution, and keep computer use.

Watch the safety prompts. When Hermes tries a remove command, it blocks it and asks for approval, showing the directories involved. Because you are handing an agent real power, run this on a machine with nothing critical on it, a spare Mac mini or Raspberry Pi is ideal. After the self-edit, the install shrank from 5.2 MB to 160 KB, about 97% smaller. Before restarting, have it set up the cron job and a memory entry so your intent survives into the next session, then restart with Control-C and typing hermes again. On the fresh session, ask it to make a plan covering an audit of the computer, practicing computer use, organization conventions, and loading on boot.

Enabling computer use and the context patch

Computer use needs the Cua driver. Hermes installs it for you, then you grant macOS permissions. Copy the permission command it gives you, paste it into the terminal, and approve the screen recording and accessibility prompts. You will likely have to quit and reopen the terminal and rerun the command, and reset the Hermes session so the computer use skill loads. Expect several permission dialogs.

There is a known bug where screenshots from the computer use tool overflow the context window. Depending on when you watch, the Hermes team may have already patched it for custom providers. To check and fix, feed Hermes the patch prompt and ask it to verify against its own code:

Check your code. Do we need this patch?

If the repository already has the fix, Hermes tells you it is good and you move on. This is the kind of issue the agent can resolve on itself, but you need to be aware it can surface unexpectedly.

Practicing with Reminders and building a widget

Start a clean conversation and practice computer use on a concrete app. The video uses the macOS Reminders app: "Let's practice the computer use tool using the Reminders app." Hermes loads the Apple Reminders skill, drives the cursor, and creates lists tied to your projects folder. A practical tip: close any windows that do not need to be open, because the agent's focus can get stolen while it learns. You can also point it at Obsidian if you want markdown flexibility instead of the default apps.

Then push further into building something native. Ask Hermes to start on boot and to create a desktop widget. It walks you through installing Xcode, which is how Apple software gets built, and generates a widget. Expect to finesse placement and behavior with follow-up prompts. In this run, the widgets could not accept text input or be moved, so the build pivoted to a dashboard showing sessions, scheduled jobs, and a button that opens the terminal with Hermes running.

Results on an open-source model are not perfect here, and a frontier model like GPT 5.5 or Opus 4.7 would likely produce cleaner output. The trade is cost and privacy against raw performance: the open-source path saves significant inference money and keeps your data private. If you are unsatisfied, you can always switch to a more expensive frontier model and accept the privacy trade-off.

Where this leaves you

The end state is a fully private AI agent living on your own computer, acting as a manager and assistant: cleaning files, organizing the desktop, maintaining reminders, and building native UI. Everything routes directly through Venice to an open-source model with zero data retention, no messaging middleman, and no creator on the other end knowing it is you.

Treat this as a starting point. Mess around, see what works and what does not, and keep it on a low-risk machine while you learn what the agent is capable of.

Key takeaways

• Connect Hermes to Venice as a custom endpoint (https://api.venice.ai/api/v1) with an API key, and consider the DIEM token for daily API credit.
• Use the Venice model table to pick something private and capable; Qwen 3.6 27B, Kimi K2.6, and DeepSeek v4 cover every tool far cheaper than frontier models.
• Run locally and skip messaging platforms to keep the agent fully private with zero data retention.
• Let Hermes optimize its own tools and skills, grant computer use permissions carefully on a low-risk machine, and check whether you need the context-overflow patch.