Getting started with DeepSee

August 19, 2026 · View on GitHub

This guide takes you from a clean machine to a verified visual conversation. Most users only need the first three sections.

Before you begin

Install Node.js 24 or newer.
Keep a working internet connection for the first installation.
Have at least one DeepSeek Harness text model for the main conversation.
For vision, prepare either a multimodal API model or enough disk space for one local OCR engine.

You do not need to install a separate DeepSee server. If dsh is not already available, the installer uses the exact Harness runtime tested with this DeepSee release.

The installer adds the same standard bundle to the Harness web and headless profiles. Each profile gets up to 15 minutes per attempt and one automatic retry. Re-running the command is safe: a profile that already contains the current version is skipped.

If the connection is unusually slow, remove DeepSee's own deadline while keeping the underlying package-manager output:

npm exec --yes --package=https://github.com/WUBING2023/deepsee/releases/download/v0.6.0-alpha.30/deepsee-0.6.0-alpha.30.tgz -- deepsee install --timeout-ms 0

Other recovery options:

npm exec --yes --package=https://github.com/WUBING2023/deepsee/releases/download/v0.6.0-alpha.30/deepsee-0.6.0-alpha.30.tgz -- deepsee install --retries 3
npm exec --yes --package=https://github.com/WUBING2023/deepsee/releases/download/v0.6.0-alpha.30/deepsee-0.6.0-alpha.30.tgz -- deepsee install --profile headless
npm exec --yes --package=https://github.com/WUBING2023/deepsee/releases/download/v0.6.0-alpha.30/deepsee-0.6.0-alpha.30.tgz -- deepsee install --force

2. Configure the first visual reader

In Harness, open Settings → Models.
Add or select the provider and model used for your normal DeepSeek conversation.
Add at least one model that declares image input, or plan to use local OCR.
Open the DeepSee sidebar panel.
Under Visual reader, choose Model and select the multimodal model, or choose OCR, select an engine, and click install.

DeepSee reads Harness provider metadata and credential references. It does not display, copy, or store the raw API key in its own configuration.

Option A: multimodal model

Use this route for natural-image understanding, charts, screenshots, and questions that need semantic visual reasoning. A model appears in the visual list only when its Harness adapter confirms image input.

Option B: local OCR

Engine	Best fit	Footprint
MinerU	Complex PDFs, papers, tables, formulas, and layout recovery	Heavy
PaddleOCR	Multilingual images, scans, and general PDFs	Medium
RapidOCR	Screenshots, receipts, simple images, and low-resource CPUs	Light

Choose an engine and click Install. A small stage progress bar and one-line comparison remain visible while DeepSee tries, as needed:

an existing compatible command for that OCR engine;
an existing uv environment;
Python + an isolated venv;
a checksum-verified portable UV archive kept inside DeepSee;
the project's official source ZIP as a final package-install fallback.

When the managed installation is ready, the same control becomes Uninstall. It removes only that engine's DeepSee-managed virtual environment, caches, downloaded models, and bootstrap files. An installation found elsewhere is shown as system-managed and cannot be removed by DeepSee.

All three defaults support CPU execution. On Windows, DeepSee automatically keeps PaddleOCR model files away from non-ASCII paths; a custom OPENDS_OCR_HOME containing non-ASCII characters fails with an explicit recovery message. Python packages and model files still require network access. A source ZIP is not a complete offline model bundle.

3. Run a smoke test

Visual route

Start a new Harness conversation.
Attach a small screenshot or photo.
Ask: Describe the image and list every visible line of text.

A successful run names the selected visual reader and passes its observation back to the base model. If Harness still says the current model has no visual capability, see Vision errors.

Workflow route

After at least two executable routes are enabled, try:

/workflow Review this small project: let one agent inspect the implementation and another check the risks, then summarize the result.

Harness should show a visible Workflow. DeepSee exposes the model directory to the main model; it does not force a complex Workflow for ordinary one-step requests.

Desktop and CLI discovery

DeepSee scans for Codex Desktop, Claude Desktop, Codex/Claude Code CLI, Kimi CLI, OpenCode, and Ollama at startup. Discovery alone does not make a runtime executable.

Codex Desktop can supply its bundled App Server. Claude Desktop is launchable, but automatic execution still requires a verified Claude Code CLI.
Codex and Claude Code can be enabled after the executable, login state, model choice, and adapter pass validation.
Kimi CLI, OpenCode, and Ollama are currently discovery-only unless a stable Harness subagent adapter is available.
A failed or logged-out CLI remains disabled instead of failing later in a Workflow.

Restart Harness after installing or signing in to a CLI so DeepSee can validate it again.

ZIP fallback

Use this route if npx github: repeatedly times out.

Download deepsee-main.zip in a browser.
Extract it and open PowerShell in the deepsee-main folder.
Run:

node .\scripts\cli.mjs install --from-folder --timeout-ms 0
node .\scripts\cli.mjs doctor
node .\scripts\cli.mjs web

This path uses the prebuilt files in the ZIP. It does not run pnpm install or compile the source. Before installing, DeepSee copies the package to $DSH_HOME\deepsee\packages\<version>, so the downloaded ZIP and extracted folder can be deleted afterward.

To start Harness later:

dsh --profile web
# If dsh is not on PATH:
npx --yes @deepseek-ai/dsh@0.1.0-rc.6 --profile web

Updates and uninstall

Opening the DeepSee panel performs a cached update check, at most once every six hours by default. When a newer version is found, click Upgrade. Checks are automatic; installation is never silent. Restart Harness when the panel shows Restart to apply.

To remove the plugin while keeping model preferences, routing state, and MinerU state:

npm exec --yes --package=https://github.com/WUBING2023/deepsee/releases/download/v0.6.0-alpha.30/deepsee-0.6.0-alpha.30.tgz -- deepsee uninstall

User state remains under $DSH_HOME/deepsee. Remove that directory manually only when you intentionally want a full reset.