dsh-tool-camera

Give DeepSeek Harness eyes — the agent captures a frame from your webcam itself, no screenshots from you.

A DeepSeek Harness plugin (bundle). After installing, the agent gains two tools:

Tool	What it does
camera_snapshot	Capture a live frame and hand the actual image to the model. Optional device / frame count / quality
camera_devices	List cameras, ffmpeg readiness, and the active capture backend

Key point: no Camera app, no foreground window, no screen capture. ffmpeg talks to the native capture backend, and pixels go straight from the device to the model.

---

The problem it solves

To let an agent see the physical world the usual trick is: open the Camera app → position the window → screenshot → feed the image. That breaks down when nobody is at the desk, when the session is remote, or when another window covers the preview.

This plugin captures a frame with one ffmpeg invocation, saves it as an attachment, and returns it as an image content block — the exact same render path as the built-in read_image.

---

Install

The package is 16 KB (ffmpeg is not bundled). Restart DSH afterwards — the tools only appear on a fresh load.

Option 1: one-shot script (recommended on Windows)

No download step needed — hand it the release URL:

pwsh -File install.ps1 -Tgz "https://github.com/zlmddygggzh-collab/dsh-tool-camera/releases/download/v1.0.1/dsh-tool-camera-1.0.1.tgz"

or point it at a tarball you already have:

pwsh -File install.ps1 -Tgz .\dsh-tool-camera-1.0.1.tgz

The script finds your profile directory, finds pnpm (PATH → corepack → the pnpm shipped with DSH → npx fallback), backs up package.json and pnpm-lock.yaml, then does both steps: pnpm add and registering the package in dsh.profile.bundles. To roll back, rename the generated *.bak-camera-<stamp> files.

With a URL it first lets pnpm fetch the tarball (so the recorded dependency stays a URL and a later pnpm install can refetch it). On hosts where pnpm's native network stack is refused at the socket layer (os error 10013), the script falls back to downloading with curl.exe into <profile>\.dsh-tool-camera\ and installing from that local file.

Option 2: two manual steps

cd ~/.dsh/profiles/desktop
pnpm add /path/to/dsh-tool-camera-1.0.0.tgz

Then append the package name to the dsh.profile.bundles array in the same package.json:

"dsh": {
  "profile": {
    "bundles": [
      "@deepseek-ai/dsh-base",
      "dshmarket",
      "...",
      "dsh-tool-camera"
    ]
  }
}

Do not skip this. bundles decides which packages are loaded into the layer stack; a dependency that is not listed there never gets mounted.

Option 3: from npm (already published)

pnpm add dsh-tool-camera

and add it to dsh.profile.bundles as above.

Uninstall

cd ~/.dsh/profiles/desktop
pnpm remove dsh-tool-camera

then drop dsh-tool-camera from dsh.profile.bundles. This package's cordis.patch.yml only inserts its own row and reconfigures nothing else, so removal is clean.

---

Usage

After restart, just ask the agent to "take a look at the camera"; it calls camera_snapshot itself.

Take a snapshot with camera_snapshot, 8 frames

Parameter	Default	Description
device	first available	Device name (see camera_devices)
frames	5	Frames captured, keep the last. Early frames are often black or over-exposed
quality	3	JPEG quality 1–31, lower is better
brightness	0	−1 to 1
gamma	1	0.1 to 10
contrast	1	0 to 3
sharpen	0	0 to 2
autoLevel	false	Stretch the frame to full dynamic range (normalize)

A dim room used to come back as a black rectangle. Now:

Take a snapshot with camera_snapshot, autoLevel on, gamma 1.6

---

Where ffmpeg comes from

The plugin does not bundle ffmpeg (60 MB would turn a two-second install into a one-minute one). On the first real capture it resolves in this order:

1. ffmpegPath from plugin config
2. DSH_FFMPEG environment variable
3. $DSH_HOME/tools/ffmpeg (ffmpeg.exe on Windows)
4. ffmpeg on PATH
5. None found → download the official prebuilt package for this platform from an npm mirror, extract the single binary into $DSH_HOME/tools/

Downloads try registry.npmmirror.com first, then registry.npmjs.org. Once in place the binary is reused forever and nothing else hits the network.

Supported: win32-x64 / win32-ia32 / darwin-x64 / darwin-arm64 / linux-x64 / linux-arm64 / linux-ia32.

Offline: set autoDownload: false, drop an ffmpeg into $DSH_HOME/tools/, or point ffmpegPath at one.

---

Configuration

Override the same row in your profile's cordis.patch.yml (no need to touch files inside the package):

- id: tool-camera
  config:
    device: ov9734_azurewave_camera   # pin one camera
    frames: 8
    quality: 2
    timeoutMs: 30000
    autoDownload: true
    gamma: 1.5                         # lift the shadows by default

Key	Default	Description
ffmpegPath	''	Explicit ffmpeg binary, highest priority
device	''	Pin a device; empty means first available
frames	5	Frames captured, max 30
quality	3	JPEG quality 1–31
brightness	0	−1 to 1; 0.2–0.35 for a dim room
gamma	1	0.1 to 10; 1.4–1.8 lifts shadows without blowing highlights
contrast	1	0 to 3; 1.2–1.4 if the lifted frame looks flat
sharpen	0	0 to 2; 0.6–1.0 to recover edges lost to noise
autoLevel	false	Stretch to full dynamic range (normalize)
timeoutMs	20000	Per-capture timeout
autoDownload	true	Fetch ffmpeg when missing

Tool arguments win over configuration, so both paths lead to the same ffmpeg filter chain.

---

Platform notes

Windows — DirectShow backend. Device names look like ov9734_azurewave_camera. If the Camera app, WeChat video, or a meeting client holds the device, ffmpeg reports Could not run graph; the plugin translates that into a readable message.

macOS — AVFoundation backend, devices are indices (0, 1, …). Grant camera access to DSH/Terminal in System Settings → Privacy & Security → Camera on first use.

Linux — V4L2 backend, devices are /dev/videoN. The current user must be in the video group.

---

Known limits

• One still frame per call, not a video stream. Call repeatedly for continuous observation.
• A camera can only be opened by one program at a time; ffmpeg holds it during capture.
• Frames are raw device output with no rotation correction — built-in laptop cameras are often flipped or mirrored depending on the driver.
• camera_snapshot declares isConcurrencySafe: () => false, so concurrent calls queue instead of fighting over the device.

---

License

MIT