Jonathon's AI Wiki

Hermes Agent Computer Use + Venice API — Private macOS Control Tutorial

9 min read

Source: Hermes_Agent_Can_Now_Manage_Your_Computer_Privately_Hermes_Agent_Tutorial.md — YouTube transcript, video, fetched 2026-05-23

A 13-minute YouTube tutorial walking through a fresh install of Hermes Agent paired with the Venice API as a zero-data-retention model provider, then activating the new Computer Use feature on macOS. The creator is unnamed in the transcript but the closing CTA points viewers to “the Venice Discord,” suggesting Venice affiliation. The walkthrough covers four distinct capabilities: (1) wiring Venice as a custom OpenAI-compatible endpoint, (2) Hermes self-cleaning its own skill bundles for context efficiency, (3) installing the Kua driver to grant macOS screen-recording + accessibility permissions, and (4) a practice run using Apple Reminders + a Xcode-built desktop widget. The privacy thesis is the through-line: open-source models on Venice are cheaper and zero-retention, with the trade-off that they’re less capable than Opus 4.7 or GPT-5.5 at agentic computer control.

Key Takeaways

  • Computer Use is brand-new in Hermes Agent. The tutorial explicitly flags it as new at recording time. Activation requires installing the Kua driver and granting both screen recording and accessibility permissions in macOS System Settings — two separate prompts, terminal must be relaunched between them.
  • Venice API integration is via custom endpoint. Venice isn’t (yet) a first-class provider in Hermes setup, so users add it as a custom OpenAI-compatible provider at https://api.venice.ai/api/v1. API key is generated from the Venice user interface; docs at docs.venice.ai.
  • Diem token (DIEM) economics. Staking the Diem token earns $1/day in recurring Venice API credit. Framed as economically meaningful for daily agent usage where token costs accumulate.
  • Open-source models on Venice are dramatically cheaper than frontier closed models. Tutorial-cited prices: Kimi K2.6 at 4.66/M output vs GPT-5.5 at 37.50/M output (~7-8× cheaper). Qwen3 27B (model #13) is even cheaper than Kimi and supports function calling + reasoning + vision + code optimization with a 256K context. DeepSeek V4 is the third recommendation. (Model names per transcript auto-caption — see Open Questions.)
  • Privacy properties vary per model. The Venice docs Models page shows per-model flags: private, anonymized, end-to-end encrypted. Not every model is private — pick deliberately.
  • Self-cleanup pattern: 5.2 MB → 160 KB system prompt (97% reduction). A first prompt of "help me clean up the tools, skills we won't use" causes Hermes to audit its built-in tool/skill bundles, then disable unused ones based on a custom role prompt (the tutorial used: “be my computer manager… personal digital janitor and organizational assistant”). Hermes makes its own recommendations on which to disable; user accepts/rejects per item.
  • Always keep skills enabled during cleanup — the skills toolset is what lets Hermes create new skills on demand for delegated tasks. Disabling it breaks the self-extending loop.
  • Terminal backend has six choices at setup. Local machine (default), isolated Docker container, cloud, Vercel, remote VPS, or Raspberry Pi. Tutorial picks local for the most-private path; recommends a dedicated machine (Mac mini, Pi) for safety when granting computer-use permissions.
  • Computer Use screenshot context-overflow bug was patched mid-tutorial. The screen-recording loop initially blew the context window. The tutorial includes a prompt-the-agent-to-patch-its-own-code workaround stored in Reminders. Hermes Agent confirmed the official repo had already patched it by the time the tutorial ran — the agent can detect and skip the manual patch.
  • Daily workflow combines three primitives: Apple Reminders skill + cron jobs for daily check-ins/to-do updates + Computer Use for desktop organization. Tutorial’s organization rules: screenshots → Pictures folder, downloads >7 days → Archives subfolder, archives in root only.
  • Desktop widgets via Xcode is theoretically possible but bounded by macOS. Hermes Agent walked the user through Xcode install + a SwiftUI widget build, but ran into a hard macOS limitation: widgets cannot accept text input and cannot be moved by the user (read-only by default). Tutorial fallback was a dashboard-style window showing sessions + scheduled jobs + a “Open Hermes terminal” button.
  • Privacy-vs-performance trade-off is explicit. Open-source models on Venice are cheaper + zero-retention; frontier closed models (Opus 4.7, GPT-5.5) are more capable at agentic Computer Use tasks. The tutorial doesn’t pretend Qwen3 27B matches Opus 4.7 — it pitches “good enough for organizational tasks, way cheaper.”

Workflow Pattern

The tutorial demonstrates a four-phase setup that maps cleanly onto how to deploy Hermes for personal computer management:

  1. Install + provider config. One-line installer script → quick setup → add Venice as custom provider → paste endpoint URL → paste API key → select model from numbered list (Qwen3 27B was #13).
  2. Skill self-cleanup. First prompt = "help me clean up the tools, skills we won't use". Second prompt = role definition ("be my computer manager…"). Hermes audits + proposes disables; user approves. Restart Hermes for context savings.
  3. Plan creation. Prompt Hermes to make a plan with explicit phases (audit current state → practice computer use → set organization conventions → ongoing workflow → boot-on-login). Hermes writes the plan to a file and walks through it.
  4. Activate Computer Use. New session → Hermes loads macos-computer-use skill → installs Kua driver if missing → user grants screen-recording + accessibility permissions → restart terminal → first computer-use task.

Hermes Agent’s transparency mode (showing every tool call) is on by default in the tutorial. The creator notes this can be disabled in config — useful when running Hermes via a messaging platform like Telegram/WhatsApp/Discord where seeing every tool call would clutter the chat.

Implementation

Tool/Service: Hermes Agent + Venice API + Kua driver (macOS Computer Use)

Setup:

  • Hermes Agent: one-line installer from the Hermes site (transcript doesn’t show the exact URL; pull from official docs)
  • Venice API: signup at venice.ai, generate API key from API section of UI, endpoint = https://api.venice.ai/api/v1
  • Kua driver: installed on demand when Computer Use first loads; requires macOS screen-recording + accessibility permissions, with a terminal relaunch between the two

Cost:

  • Qwen3 27B (Venice): cheapest of the three recommended; exact pricing not stated in transcript but described as “even cheaper than Kimi”
  • Kimi K2.6 (Venice): 4.66/M output, 256K context, function calling + reasoning + vision + code optimized
  • DeepSeek V4 (Venice): third recommendation, pricing not stated
  • GPT-5.5 (reference comparison): 37.50/M output
  • Diem token staking: nets $1/day Venice API credit

Integration notes:

  • Venice models route through a standard OpenAI-compatible interface — any Hermes provider slot that accepts a custom endpoint works
  • The 256K context limit on Kimi K2.6 is meaningful for Computer Use, which generates lots of screenshot tokens per loop
  • Per-model privacy flags (private, anonymized, end-to-end encrypted) are visible in the Venice docs Models page — verify before assuming any given model is zero-retention
  • The skill-cleanup pass takes the system prompt from ~5.2 MB to ~160 KB on a fresh install; rerun after any major Hermes update that may add back default skills
  • Computer Use on macOS requires Kua driver permissions to be re-granted any time the terminal is reinstalled or moved

Try It

  1. Install Hermes Agent from the official Hermes installer. On macOS this is one terminal command; let it provision Python/Git/Node and run hermes setup. Pick “quick setup” when prompted.
  2. Add Venice as a custom provider. Endpoint URL: https://api.venice.ai/api/v1. Sign up at venice.ai → API section → “Generate new API key” → label it “Hermes Agent” → paste into the Hermes setup prompt.
  3. Pick a model with care for privacy properties. Open docs.venice.ai → Models → Text and scan the per-model flags (private, anonymized, end-to-end encrypted). For Computer Use you want function calling + vision + reasoning + code optimized — Kimi K2.6 or Qwen3 27B both meet this bar. Note the model number from the Hermes-side list when prompted.
  4. Select a terminal backend. Local for maximum privacy, Docker for isolation, Vercel/VPS/Raspberry Pi if running Hermes off your personal machine. Local is the default — pick deliberately if not.
  5. Run the skill-cleanup pass on first launch. First prompt: help me clean up the tools, skills we won't use. Then describe your role for Hermes (e.g., be my computer manager — helping me stay organized with files, desktop, etc.). Approve the proposed disables but keep skills and code execution enabled.
  6. Activate Computer Use. New conversation → prompt: activate the computer use tool and make a plan for setup. Install Kua driver when prompted. Grant macOS screen-recording permission (System Settings → Privacy & Security → Screen Recording → Terminal). Relaunch terminal. Grant accessibility permission. Relaunch terminal again. Hermes will retry the Computer Use call.
  7. Stake DIEM (optional) for ongoing Venice API credit if planning daily agent use — $1/day adds up against API spend.
  8. Run on a sacrificial machine. The transcript repeatedly warns: granting AI screen-recording + accessibility + file-system access to a primary-work machine is a real risk. Mac mini or Raspberry Pi recommended.

Open Questions

  • Creator identity / Venice affiliation status. The transcript closes with “Share with us what you’re working on in the Venice Discord. There’s a link below.” The creator is not named anywhere in the transcript. This is most likely an official Venice-affiliated channel, but unverified from the transcript alone — the video page or channel branding would confirm.^[inferred]
  • Whether the screenshot context-overflow patch is in main Hermes or a custom branch. The tutorial says Hermes Agent confirmed the patch was already in the official repo at recording time, but doesn’t cite the commit. Verify against Hermes Agent changelog when documenting this for users on older releases.
  • Whether the Xcode widget workaround is documented anywhere upstream. macOS widget input limitations (read-only by default) appear to be a hard platform constraint, not a Hermes issue. The dashboard-style window with a “Open Hermes terminal” button was the tutorial’s fallback — whether anyone has documented a cleaner widget-or-companion-app pattern for Hermes Agent on macOS is unknown.
  • Model name normalization. Transcript auto-captions render model names imperfectly: “Qwen 3.6 27B” likely = Qwen3 27B; “Kimi Cape 2.6” almost certainly = Kimi K2.6; “DeepSeek version four” / “DeepSeek v4” = DeepSeek V4; “Clodopus 4.7” = Claude Opus 4.7; “GPT 5.5” assumed = GPT-5.5 (uncertain). “Kua driver” may be the upstream Cua driver (computer-use abstraction layer) under a transcribed spelling. Verify model names against the live Venice docs Models page before quoting prices to anyone.

Graph View

Backlinks