Vision: AI That Actually Sees Your Images

Imagine sending a screenshot to a friend and saying, “What do you notice here?” — except the friend is your AI agent, and it answers instantly. That’s exactly what Hermes Agent can do now.

What Does This Mean?

Hermes Agent can look at images you paste into the terminal. You don’t need to upload files or type file paths. Just copy an image, attach it, and ask questions like “What’s wrong with this error message?” or “Describe this chart.”

How to Attach an Image (3 Simple Ways)

Method 1: The /paste Command (Most Reliable)

Type /paste and press Enter. Hermes checks your clipboard and attaches whatever image you copied.

/paste

This works almost everywhere — even when your terminal normally messes up Cmd+V or Ctrl+V.

Method 2: Just Press Cmd+V / Ctrl+V

If you’re on macOS Terminal, iTerm2, or most Linux desktops, you can simply paste like you normally would. Hermes is smart enough to detect when your clipboard contains an image and attach it automatically. Pasted macOS screenshot temp paths and file://... image URIs also attach right away instead of sitting in the composer as raw text.

Method 3: /terminal-setup for VS Code Users

If you use VS Code, Cursor, or Windsurf as your terminal, run this once:

/terminal-setup

This fixes common keyboard shortcuts like Cmd+Enter and undo/redo that the IDE sometimes intercepts. Run it on your local machine only — not inside an SSH session.

What Happens After You Paste?

Once you attach an image, you’ll see a small badge above your input:

[📎 Image #1]

You can attach multiple images before sending — each gets its own badge. Press Ctrl+C to clear them all. Then type your question and press Enter. The image is sent to the AI model as a vision block, and it processes it just like a human would look at a picture. Images are saved to ~/.hermes/images/ as timestamped PNG files.

Which Platforms Work Best?

Environment	Paste Support	Notes
macOS Terminal / iTerm2	✅	Best experience
Linux (X11)	✅	Needs xclip installed
Linux (Wayland)	✅	Needs wl-paste installed
WSL2 (Windows Terminal)	✅	Works out of the box
VS Code / Cursor / Windsurf	✅	Run /terminal-setup first

A Quick Tip for Clipboard-Only Images

Here’s a common gotcha: if your clipboard has only an image (no text at all), some terminals can’t send the image bytes directly. That’s why /paste exists — it’s your safety net.

Pro tip: If you’re on macOS and take a screenshot, Hermes can even recover the screenshot’s temporary file path automatically. Just paste and go.

Follow-Up Questions About an Earlier Image

The session database stores user messages as text, with one [screenshot] line per attached image. When a turn’s history is rebuilt from the database (an API server request continuing a session with X-Hermes-Session-Id, hermes chat --resume, or a gateway restart), Hermes now re-sends the most recent images in their original message — but only when the model takes images natively. Control it in config.yaml:

vision:
  replay_recent_images: 3   # default; 0 = never re-send earlier images

The bytes come from the copy Hermes already keeps under cache/documents/native-inputs/ in the profile, so nothing new is stored and the database stays text only. A re-sent message is byte-identical to the original request, so the provider’s prompt cache stays valid. With the default 3, the newest 3 to 5 images are re-sent, and every third new image drops the oldest batch at once. Replay only reaches images whose kept copy still exists: on the gateway, a chat image’s kept copy is removed after 24 hours, so a follow-up more than a day later sees [screenshot] again. Images outside the window, remote URLs, and missing or changed files also keep the [screenshot] text. With agent.image_input_mode: text or a model without native vision, nothing is re-sent, and a value that isn’t a non-negative integer logs a warning and disables the replay.

What If Your Model Can’t See Images?

Not every model supports vision. Hermes handles this for you: if your current model is vision-capable, the image is sent as real pixels. If it’s text-only, Hermes routes the image through a helper vision model that describes it in words, then feeds that description into the conversation. You don’t have to configure anything — switch between models mid-session and image handling just works.

If you need to force a specific behavior, set agent.image_input_mode in config.yaml:

• auto (default) — native pixels when the model supports vision, text description otherwise.
• native — always attach pixels, even for text-only models.
• text — always use the text-description path, never attach pixels.

This is handy when a backend accepts text but rejects native image input. Point auxiliary.vision at a different vision-capable provider and model, and images switch to the description path automatically. The same decision applies to every other tool that hands the model an image — browser screenshots, computer_use captures (described by the auxiliary vision model when the main model can’t take them natively), and MCP image results. So agent.image_input_mode: text keeps pixels out of all of them.

Summary

Hermes Agent now lets you paste images and get AI-powered analysis — no file uploads, no complex setup. Copy an image, use /paste (or just paste), ask your question, and watch the magic happen.

Practical tip: Start with /paste when you’re unsure. It’s the most reliable method across all platforms. Once you get comfortable, try plain Cmd+V — it works great on most systems and feels even more natural.