生成媒体CLI

Install

For Linux and macOS:

curl https://genmedia.sh/install -fsS | bash

For Windows:

irm https://genmedia.sh/install.ps1 | iex

Setup

genmedia setup

Interactive wizard that configures:

• API key — saved encrypted to your local config (or skip and use FAL_KEY in your environment)
• Auto-load .env — automatically load FAL_KEY and related vars from a project .env file
• Output mode — auto (pretty in TTY, JSON when piped), json (always structured), or standard (always human-readable)
• Automatic updates — check for new versions in the background and swap in on next launch (default: on; set GENMEDIA_NO_UPDATE=1 to disable)

Get your API key at fal.ai/dashboard/keys.

To skip the wizard, set the key in your environment:

export FAL_KEY=your_fal_api_key

Non-interactive setup (agents / CI)

genmedia setup --non-interactive --api-key "$FAL_KEY"
genmedia setup --non-interactive --output-format json --no-auto-load-env --auto-update

Every flag is optional — fields you don't pass keep their current values, so repeated invocations are idempotent.

`init` — Install the default genmedia skill bundle

genmedia init
genmedia init --force   # overwrite existing files

Installs the default genmedia skill bundle:

The bundle is installed into .agents/skills/ if the project has a .agents/ directory, otherwise into .claude/skills/. If neither directory exists, the command exits with a message asking you to create one and try again.

After running init, compatible agent sessions in that project can use the installed skills without needing to call --help. Commit the installed skills directory so teammates and other agents get the same context.

`skills` — Manage installed agent skills

genmedia skills list                  # list everything in the registry
genmedia skills list image            # full-text search name + description
genmedia skills install genmedia
genmedia skills update
genmedia skills remove genmedia

Installs, updates, lists, and removes agent skills from the genmedia registry. skills list <query> runs full-text search via the hosted index API (GENMEDIA_SKILLS_API_URL, default https://genmedia.sh/skills); the no-query form fetches index.json directly from the registry.

`gallery` — Browse a per-session HTML preview of generated assets

Every successful genmedia run (and genmedia status --download) records its outputs into a per-session, self-contained static HTML page under ~/.genmedia/gallery/sessions/<session_id>/index.html. The page aggregates every image / video / audio / 3D asset produced inside one agent session — across many runs and many endpoints — into a filterable grid with a lightbox view. No server, no dependencies, opens straight from file://.

genmedia gallery                          # pretty TTY: prints `→ Open: <url>`. JSON / non-TTY: full info payload.
genmedia gallery --json                   # structured info payload, agent-safe
genmedia gallery info                     # full info payload, even from a TTY (devs)
genmedia gallery open                     # open the all-sessions index in the default browser
genmedia gallery open current             # open the current session
genmedia gallery open latest              # open the most-recently recorded session (works across shells)
genmedia gallery open <session_id>        # open a specific session
genmedia gallery open latest --print      # resolve path/url without launching the browser
genmedia gallery list                     # list every recorded session (newest first, JSON-friendly)
genmedia gallery list --limit 10          # cap the list (default: 50)
genmedia gallery rename --label "demo run"        # rename the current session (cosmetic — id stays the same)
genmedia gallery rename latest --label "x"        # rename the most-recent session
genmedia gallery rename <session_id> --clear      # remove a session's label
genmedia gallery clear --yes              # delete the current session (--yes is required)
genmedia gallery clear latest --yes       # delete the most-recently recorded session
genmedia gallery clear <session_id> --yes # delete a specific session
genmedia gallery clear all --yes          # delete every recorded session

The page auto-refreshes via a tiny client-side poll on ./data.json while a session is "live" (updated in the last 5 min), so new runs land in the open tab without a manual reload.

gallery clear always requires --yes and never prompts interactively — same shape for agents and humans. GENMEDIA_NO_GALLERY=1 only disables recording; list / open / clear still work so you can inspect or purge prior galleries.

The run and status --download responses include the gallery info too:

{
  "status": "completed",
  "endpoint_id": "fal-ai/flux/dev",
  "request_id": "…",
  "gallery": {
    "session_id": "a1b2c3d4e5f6",
    "path": "/Users/you/.genmedia/gallery/sessions/a1b2c3d4e5f6/index.html",
    "url": "file:///Users/you/.genmedia/gallery/sessions/a1b2c3d4e5f6/index.html"
  }
}

Session detection (agent-agnostic). The session id is resolved in this order, so the gallery groups outputs correctly across Claude Code, Codex, Cursor, Amp, Gemini, Aider, and anything else:

1. GENMEDIA_SESSION_ID — explicit override (e.g. set from a hook script).
2. Per-agent env vars: CODEX_THREAD_ID, CLAUDE_SESSION_ID, AMP_THREAD_ID, CURSOR_TRACE_ID.
3. Process-tree walk — looks for a known agent binary (claude, codex, cursor, amp, gemini, copilot, opencode, aider, cline) in the ancestor chain via ps -e. Critical for Claude Code, whose Bash tool spawns a fresh shell per call: process.ppid changes every invocation, but the Claude process itself lives for the whole conversation and becomes the stable anchor.
4. Terminal-level: TERM_SESSION_ID / ITERM_SESSION_ID / WT_SESSION / TMUX_PANE.
5. Fallback: a deterministic hash of <agent>:<ppid>.

or persist in config:

genmedia setup # (toggle is currently file-only — set "analyticsOptOut": true) ```

Builds from source without a POSTHOG_KEY environment variable (the open-source default) report nothing — the analytics module no-ops at runtime.