Realsee Skills — Argus Agent and CLI Workflow

English | 简体中文

Realsee Argus is a world-leading 3D vision foundation model. From a photo, panorama, or sparse views, it reconstructs metric 3D structure in milliseconds: pose, depth, point clouds, and renderable geometry.

Realsee Skills provides the Argus agent and CLI workflow. The current Skill is argus 2.0: it processes 1–99 exact 2:1 panoramas and produces EXR depth maps, one merged GLB point cloud, per-image camera poses, optional intrinsics, and a validated local result index.

The Skill ID remains argus. Version 2.0 has no legacy single-image VGGT fallback. Pin v1.0.2 when a workflow needs square 1:1 input, the old single-GLB-only result, or the old H5 preview behavior.

Argus · Interactive demo · Research · Developer Platform

Official evidence includes a 1.31B-parameter model and the Realsee3D benchmark's 10K complete indoor scenes, 95,962 room units, and 299,073 panoramic viewpoints. These are model and benchmark figures, not the Skill input limit.

The installable Skill 2.0 contract remains intentionally specific: 1–99 local RGB8 panoramas with exact 2:1 dimensions. Capabilities shown for photos, sparse views, or other product surfaces are not exposed by this CLI.

Credentials

Every install path uses the unchanged runtime contract:

Key	Purpose	Sensitive
REALSEE_APP_KEY	Realsee Open Platform APP_KEY	yes
REALSEE_APP_SECRET	Realsee Open Platform APP_SECRET	yes
REALSEE_REGION	global (app-gateway.realsee.ai) or cn (app-gateway.realsee.cn)	no

Register at my.realsee.ai or my.realsee.cn, then request the Argus Gateway capability through the support channel described in SUPPORT.md.

Install

Claude Code marketplace:

/plugin marketplace add realsee-developer/skills
/plugin install realsee-skills@realsee-developer-skills

Codex:

npx skills add realsee-developer/skills --skill argus --agent codex

Any detected agent host:

npx skills add realsee-developer/skills --skill argus
npx skills add realsee-developer/skills --skill argus --agent '*'

Install from a local checkout:

git clone https://github.com/realsee-developer/skills.git
cd skills
npm install
(cd .agents/skills/argus && npm ci --omit=dev --ignore-scripts --no-audit --no-fund)
npm run rebuild

See the install overview, Claude Code, and Codex for host-specific details.

Official example manifest

Every Skill install includes examples/manifest.json, which lists the CDN URL, byte length, and SHA-256 for the CN and Global first-party sample sets. Panorama JPEGs are absent from the current release tree and every generated Skill distribution. Download one set to a new absolute directory outside the installed Skill:

node <skillDir>/scripts/download-examples.mjs \
  --region cn \
  --output /absolute/example-output

Use the set matching REALSEE_REGION. The downloader publishes the output only after every file passes its manifest checks. Running Argus is a separate remote upload and still requires user consent.

Direct CLI

Start from repeated images:

node .agents/skills/argus/scripts/run-argus.mjs start \
  --image /absolute/path/a.jpg \
  --image /absolute/path/b.webp \
  --workspace /absolute/workspace-root \
  --yes --json

Or start from one existing ZIP:

node .agents/skills/argus/scripts/run-argus.mjs start \
  --zip /absolute/path/input.zip \
  --workspace /absolute/workspace-root \
  --yes --json

Capture the returned workspace_dir. Each status call makes one query:

node .agents/skills/argus/scripts/run-argus.mjs status \
  --workspace /absolute/workspace-root/<run-dir> --json

Collect after success:

node .agents/skills/argus/scripts/run-argus.mjs collect \
  --workspace /absolute/workspace-root/<run-dir> --json

There is no detached poller, --async, or --resume. Completed collection is idempotent.

Input and output

Inputs are 1–99 JPEG, PNG, or WebP RGB8 panoramas with exact width == 2 * height. At least 2048×1024 is recommended; lower resolution is a warning. --image is repeatable and mutually exclusive with --zip. ZIP mode is safely extracted, validated, normalized to Unicode NFC, sorted, and repacked before upload.

Remote task_status (queued, processing, succeeded, failed) is separate from algorithm result_status (success, partial, error). A partial result exits 0 but includes a prominent warning and non-empty missing_ids; an error exits non-zero.

The collector retains output.zip, safely extracts it, and writes this validated local result matrix:

Artifact	Availability
output.json	Required algorithm manifest.
pointcloud/merged.glb	One merged right-handed, Y-up point cloud for successfully reconstructed images.
depth/*_depth.exr	One meter-scale floating-point depth map per successful image.
pose/*_pose.json	One camera pose per successful image.
intrinsics/*_intrinsics.json	Optional; absence is valid.
result.json	Local index for statuses, paths, warnings, and missing_ids.

Real Argus runs upload the normalized input ZIP to Realsee remote services. Obtain user consent before upload. Never commit or log credentials, upload tokens, private result URLs, or generated artifacts.

Contracts and migration

• Skill README
• Official brand asset manifest
• Gateway OpenAPI
• Algorithm I/O contract
• output.json JSON Schema
• Migration from 1.x
• Machine-readable index

Development

Canonical source lives in .agents/skills/argus/. The Claude plugin and CN-only Arkclaw package are generated from it and checked for byte consistency (with deterministic Arkclaw overlays for runtime region, example downloads, and matching guidance).

npm run doctor
npm run test:skill
npm run rebuild
npm run ci

See Architecture, development, release, and public distribution.

License

This repository is source-available under the Realsee SDK License Agreement. It is not published under an OSI open source license.