webcam-cli

Agent and CLI for USB webcam and microphone access: enumerate attached devices, capture stills and video clips, and record audio, with agent-friendly verbs and safe-by-default writes.

Status

Scaffold. The agent-first CLI skeleton, mesh identity, CI, and skill kit are in place and green. The capture surface — list, describe, capture, record — is not implemented yet. The design brief and the questions still to settle live in issue #1.

Scope

webcam-cli owns local USB capture devices — video and audio input — and the act of getting frames and samples off them. It produces an artifact plus honest metadata about it, and stops there.

It does not own interpreting what is in a frame (a vision model's job), sound out (harmonics-cli), browser or screen capture (webglass-cli), the general file-and-shell surface (shell-cli), or mesh event semantics (events-cli).

Quickstart

uv sync
uv run pytest -n auto                 # run the test suite
uv run webcam whoami                  # identity from culture.yaml
uv run webcam learn                   # self-teaching prompt (add --json)
uv run teken cli doctor . --strict    # the agent-first rubric gate CI runs

The console command is webcam. The import package is webcam_cli and the PyPI distribution is webcam-cli — deliberately decoupled, so the ergonomic thing you type stays short while the import cannot shadow a generic webcam module in a consumer's environment.

Known defect: the CLI does not yet call itself by its installed name. webcam --help prints usage: webcam-cli, and error hints, learn, and explain all tell you to run webcam-cli … — which is not installed and fails with Failed to spawn. Wherever the CLI says webcam-cli, type webcam. Tracked in issue #3; the fix lands with the self-description rewrite, since the same strings are also still describing this repo as a template.

CLI

Today's surface is the agent-first template baseline:

Planned, per the build brief:

Every command supports --json. Results go to stdout, errors/diagnostics to stderr (never mixed). Exit codes: 0 success, 1 user error, 2 environment error, 3+ reserved.

Why device identity is the hard part

/dev/videoN numbering is plug order, not identity. On the operator's host the two attached cameras swapped indices between the brief being written and being re-checked — the C270 moved from video2 to video0 and the Arducam the other way, with their ALSA cards swapping to match. A UVC camera also commonly exposes two /dev/video* nodes, only one of which yields frames, so counting nodes overcounts cameras. Stable handles come from /dev/v4l/by-id/, which carries vendor, product, and serial.

Access is the other trap: on that host the operator is not in the video group at all — read/write comes from a seat ACL granted by logind. An agent running headless, in a container, or as a systemd unit will not receive it and will fail to open a camera that works fine from a desktop session.

See issue #1 for the full evidence and the open questions it raises.

What this repo carries

• An agent-first CLI cited from teken (afi-cli) — the runtime package has no third-party dependencies.
• A mesh identity — culture.yaml (suffix + backend) and the matching resident prompt file (AGENTS.colleague.md, since this agent runs backend: colleague).
• The canonical guildmaster skill kit under .claude/skills/, vendored cite-don't-import. See docs/skill-sources.md.
• A build + deploy baseline — pytest, lint, the agent-first rubric gate, and PyPI Trusted Publishing wired into GitHub Actions.

See CLAUDE.md for the architecture, the domain constraints, and the conventions (version-bump-every-PR, the cicd PR lane, worktree layout, deploy setup).

License

Apache 2.0 — see LICENSE.