# Hermes Agent vision: send it screenshots and images > How to attach images in the terminal or from chat apps, how Hermes routes them for vision and text-only models, and the settings that keep image-heavy sessions affordable. Canonical: https://hermeswatcher.com/learn/vision-and-images Section: Learn ยท Skills, plugins & MCP Last reviewed: Oct 20, 2026 Basis: Explained from the official Hermes Agent docs and repository (Nous Research). Steps follow the linked official sources. ## Getting an image into the conversation In the CLI, copy an image and either paste it or type `/paste`. Each attached image shows up as a badge above the input. You can attach several before sending, and `Ctrl+C` clears them. Terminals can't send raw image bytes, so if your clipboard holds only an image and pasting does nothing, use `/paste`, which reads the clipboard directly. Attached images are saved as PNG files under `~/.hermes/images/`. From your phone, just send the photo to your bot on Telegram, Discord, Slack or WhatsApp. These platforms handle the upload themselves, so none of the clipboard limitations below apply. You can also paste an image URL into a message, and the agent can look at it with the `vision_analyze` tool. ## Clipboard setup by platform - macOS: works out of the box using `osascript`. Installing `pngpaste` makes it faster. - Linux on X11: install `xclip`. On Wayland: install `wl-clipboard`. Run `echo $XDG_SESSION_TYPE` to find out which one you're using. - WSL2: no install needed. Hermes reads the Windows clipboard through `powershell.exe`. - Over SSH: clipboard images don't work, because the clipboard tools read the remote machine's clipboard, not yours. Copy the file over with `scp` and reference it by path, use a URL, or send the image through a messaging app. ## Vision models and text-only models You don't need to choose a special model for images to work. If the current model supports vision, Hermes sends the actual pixels. If it doesn't, the image goes through `vision_analyze`, where an auxiliary vision model describes it and the description is added to the conversation. This happens automatically, so you can switch models mid-session without changing how you work. Browser screenshots, computer-use captures and MCP image results all follow the same rule. To override the automatic choice, set `agent.image_input_mode` to `native` (always send pixels), `text` (always send a description) or leave it on `auto`. The documented reason to change it is a backend that accepts text but rejects image input. In that case, keep your main model and point `auxiliary.vision` at a different vision-capable provider and model. In `auto` mode, an explicitly configured `auxiliary.vision` backend already switches images to the description path. ## Keeping image-heavy sessions affordable An image embedded natively is re-sent with every later request in the session, so it keeps costing tokens. Two settings limit that. `vision.embed_target_bytes` (default 256 KB) downscales large images to fit. Raise it, for example to `1048576`, if dense screenshots of tables come back unreadable. `vision.max_calls_per_image` caps how often the same image can be embedded. By default only delegated subagents are capped, at 3, but you can set a number for every session. `vision.replay_recent_images` (default 3) re-sends your most recent images when a session is rebuilt from history, for example after a resume or a gateway restart, so a follow-up question about a photo still sees it. Set it to `0` to turn that off. On the gateway, the stored copy of a chat image is cleaned up after 24 hours, so a question about a photo from yesterday may only see a `[screenshot]` placeholder. Send the image again if it matters. ## A quick checklist when an image is ignored - No badge appeared: use `/paste`, and install `xclip` or `wl-clipboard` on Linux. - Working over SSH: transfer the file, or use a messaging app instead. - The model says it can't see images: check `agent.image_input_mode` and your `auxiliary.vision` settings. - The model only sees an `[image omitted ...]` note for an SVG, BMP or TIFF on a Responses-style backend such as `openai-codex`: those backends accept only JPEG, PNG, GIF and WebP. Hermes converts SVGs to PNG when a rasterizer such as `rsvg-convert` or `cairosvg` is installed. Other formats need converting before you send them. ## Related guides - [How to choose a model for Hermes Agent](https://hermeswatcher.com/learn/models-and-effective-configuration.md) - [How to keep Hermes Agent costs down](https://hermeswatcher.com/learn/keeping-costs-under-control.md) - [Hermes Agent messaging gateway explained](https://hermeswatcher.com/learn/messaging-gateway-explained.md) ## Related tools - [Collect content ideas without losing the evidence](https://hermeswatcher.com/workflows/content-research-queue.md) - [Models and providers: configure the actual inference route](https://hermeswatcher.com/tools/models.md) - [Telegram: reach Hermes from a private bot conversation](https://hermeswatcher.com/tools/telegram.md) - [Browser tools: interact with pages when retrieval is insufficient](https://hermeswatcher.com/tools/browser-tools.md) - [Image generation: create images from a text prompt](https://hermeswatcher.com/tools/image-generation.md) - [CLI: run bounded tasks and inspect the working environment](https://hermeswatcher.com/tools/cli.md) ## Sources - [Vision and Image Paste (official docs)](https://hermes-agent.nousresearch.com/docs/user-guide/features/vision) - [Configuration: auxiliary models (official docs)](https://hermes-agent.nousresearch.com/docs/user-guide/configuration) --- HermesWatcher is an independent guide. It is not affiliated with or endorsed by Nous Research. The [official Hermes Agent documentation](https://hermes-agent.nousresearch.com/docs) and [repository](https://github.com/NousResearch/hermes-agent) from Nous Research are the authority; each page here lists the official sources it was written from. If this site and the official docs disagree, trust the official docs. Index for agents: https://hermeswatcher.com/llms.txt