Game assets with fal

fal runs hundreds of generative models behind one API key. Use it whenever the mod needs something that doesn't exist yet: a weapon sprite, a boss, a unit rendered from 16 angles, a tileable floor, a laser sound, boss music, a voiced line.

Setup (check once per session)

• Key. Create one at https://fal.ai/dashboard/keys. um fal takes FAL_KEY from the environment first, then the key the human stored with um fal key set (a hidden prompt: the desktop keyring on Linux, through libsecret's secret-tool; an owner-only ~/.secrets/fal-key where there's no secret-tool), then FAL_KEY=... in a .env in the working folder. um fal key status shows which source is used, with keys masked. Never write the key into mod files or configs; um publish check flags leaked keys.

• MCP. fal's hosted MCP server is https://mcp.fal.ai/mcp with header Authorization: Bearer $FAL_KEY. The repo pre-configures it for each agent:

  • Claude Code: .mcp.json
  • Codex: .codex/config.toml
  • Cursor: .cursor/mcp.json
  • VS Code/Copilot: .vscode/mcp.json
  • Gemini CLI: gemini-extension.json
  • OpenCode: opencode.json

  They read FAL_KEY from the agent's environment, and without it the server doesn't connect. If the fal MCP tools are missing, use um fal meanwhile and tell the human. They can start the agent with FAL_KEY set (on Linux, from the keyring: export FAL_KEY="$(secret-tool lookup service fal.ai key api)"), or add fal's OAuth endpoint, which takes a browser sign-in instead of a key:

  • Claude Code: claude mcp add --transport http fal-oauth https://mcp.fal.ai/mcp-relay, then Authenticate it in /mcp
  • Codex: codex mcp add fal-oauth --url https://mcp.fal.ai/mcp-relay, then codex mcp login fal-oauth
  • anything else: a Streamable HTTP server at https://mcp.fal.ai/mcp-relay with OAuth and no header (https://fal.ai/docs/documentation/setting-up/mcp)

  Outside a clone, with a key: Codex takes [mcp_servers.fal] with url = "https://mcp.fal.ai/mcp" and bearer_token_env_var = "FAL_KEY" in ~/.codex/config.toml. MCP tools: search_models, recommend_model, get_model_schema, get_pricing, run_model, submit_job/check_job/get_job_result, upload_file, search_docs.

• No fal key? For images, um comfy uses a local ComfyUI server instead (um comfy status, then um comfy image "..." --sprite). It has no audio, 3D or video recipes.

• Sprixen is another hosted option for 2D art only. It's a separate paid service (an account, an API key and credits), so ask the user before using it. um sprixen keeps one art style, resolution and view per project (um sprixen sprite "..." --style pixel_art --size 32x32), turns a sprite id into a one-row sheet of 2-12 frames (um sprixen animate <sprite_id> --type walk), and makes terrain tiles. The key goes in SPRIXEN_API_KEY or um sprixen key set, and um sprixen credits shows the balance. No audio, 3D or video.

• Voidless is another separate paid service for 2D art (an account on a plan with API access, an API key and credits), so ask the user before using it. um voidless sprite "..." prints an asset id; um voidless rotate <asset_id> adds the eight facings, and um voidless animate <asset_id> --preset <key> adds a set of clips at the asset's size (um voidless presets lists the keys). um voidless fetch <asset_id> downloads the frames again without generating. The key goes in VOIDLESS_API_KEY or um voidless key set, and um voidless credits shows the balance. For 8 or 16 headings from a 3D model instead, see um render3d.

• CLI alternatives: pip install fal gives fal api <endpoint> key=value key:=json; the genmedia CLI (genmedia run ... --json --download) is agent-friendly too.

Which interface

• Discovery (what's the best model for X right now, its inputs, its price): fal MCP recommend_model / search_models / get_model_schema / get_pricing, or um fal search, um fal schema <endpoint>, um fal price <endpoint>. The catalog changes weekly. The defaults below were current in September 2026; check before a big batch.
• Anything that must land on disk (every game asset): um fal <recipe>. It uploads local inputs, queues, polls, downloads every output file, and appends the endpoint, inputs, seed and request id to <out>/fal_manifest.jsonl, so every asset can be traced and regenerated.
• Quick look or one-off where a URL is enough: MCP run_model.

Recipes (um fal <recipe> --help for options; --model overrides the endpoint; --set k=v passes extra inputs)

Other useful endpoints:

• 3D: tripo3d/tripo/remesh and fal-ai/meshy/v5/remesh (low-poly, game-ready), fal-ai/trellis-2/retexture.
• Motion: fal-ai/hunyuan-motion (text → motion).
• Icons: fal-ai/recraft/v4.1/text-to-vector (SVG icons).
• Video cutouts: pixelcut/video-background-removal.
• Music: google/lyria-3.5.
• Voice: fal-ai/minimax/speech-2.8-hd.

Prompting game art that fits the game

• Look at the game's own assets first: pixel size, outline, palette, perspective, facing, how busy they are. Put that into a reusable style suffix. For Terraria: "16-bit pixel art game sprite in the style of Terraria, crisp dark outline, limited palette, centered, plain flat white background, no shadow, no text".
• Describe the view and orientation explicitly: "perfectly horizontal side view with the muzzle pointing right" for held weapons, "seen from the side facing left" for enemies, "pointing straight down" for a falling bomb. Engines have conventions (Terraria items point right, NPCs face left) and fixing orientation afterwards costs quality.
• Backgrounds: transparent (gpt-image-2) or a flat colour that um sprite cutout can flood-fill. Avoid gradients, scenery and ground shadows. If a soft shadow sneaks in, use --grey / --keep-top in cutout.
• Player / team colour: ask for "bright saturated blue accents" on the parts that should take the player's colour, then um sprite team-mask --hue blue turns them into a mask.
• Consistency across a set: generate one hero image, then derive the rest with the edit endpoint and the hero as --ref ("same robot, now firing, muzzle flash"). Don't ask one prompt for a whole sprite sheet; grids come out uneven.
• Many angles or frames of the same object: go 3D. Take the concept, run um fal model3d, then um render3d from the game's camera (asset-pipeline skill). That's how the AoE2 robotaxi got 16 consistent headings × 5 animations.
• No text or logos in art unless wanted: models love to add them.

Audio for engines

• The SFX and music endpoints return MP3. Convert to what the engine wants:
  • ffmpeg -i x.mp3 -ar 44100 x.wav (XNA/tModLoader, most engines);
  • ffmpeg -i x.mp3 -c:a libvorbis -q:a 5 x.ogg (Minecraft, Godot, Unity);
  • trim silence first: -af silenceremove=start_periods=1:start_threshold=-50dB.
• Loops: um fal sfx ... --loop, or ask the music model for a loopable track and crossfade the ends.
• Keep SFX short (0.2-2 s) and normalize loudness (-af loudnorm=I=-16) so they sit with the game's own sounds.

Cost and etiquette

• Price the expensive stuff: before batching 3D, video or long music, check the price (um fal price <endpoint> / MCP get_pricing). For more than about 20 generations or anything video-sized, tell the user the rough cost first.
• Iterate cheap: use low quality or resolution while exploring (--quality low, --res 0.5K), then re-run the winners at full quality with the same prompt (and seed where supported).
• Reproducibility: keep fal_manifest.jsonl with the assets; it records prompts, seeds and request ids.
• Credits: in the mod's README, credit that assets were generated with fal and name the models. Check a model's license page for commercial use.