SFXMint MCP Server
Free CC0 sound effects — describe a sound to find it, preview, and download WAV/MP3. No signup, no API key, no attribution. Agents can ask by role, take one of 18 sets, or search the library, and get permanent hotlinkable URLs. When nothing in the library fits, search answers with one line for the user and a short link that opens the description in AI Studio, free to make and hear there. With an API key the agent can also generate the sound itself: the account's 1 free try first, then credits or a plan.
Downloads have a daily allowance: without a key, 20 different sounds a day per network (a second format or a retry of the same sound is not counted again); with a free account's API key, 50 a day; any purchase removes the limit for 30 days, and file URLs returned to a keyed request carry a signed g= good for 48 hours. Past it a file answers 429 download_limit with a message to pass on to the user. Listening in a browser is never counted. Keep page_url in SOUNDS.md, never a URL with g=.
Sounds are AI-generated or procedurally synthesized and offered under CC0. Audio URLs are immutable, CORS-enabled and free to hotlink. Download files into the project when offline playback is required; preview suitability in the actual app or game.
Quick start
Claude Code, as a plugin (the MCP server plus the skill below):
/plugin install sfxmint --marketplace flreey/sfxmint-mcp
On Claude Code before 2.1.275, run /plugin marketplace add flreey/sfxmint-mcp, then /plugin install sfxmint@sfxmint. Or add only the MCP server:
claude mcp add --transport http sfxmint https://sfxmint.com/mcp
Gemini CLI, as an extension (the MCP server, a context file and the skill):
gemini extensions install https://github.com/flreey/sfxmint-mcp
Codex CLI, as a plugin (the MCP server plus the skill):
codex plugin marketplace add flreey/sfxmint-mcp
codex plugin add sfxmint@sfxmint
Cursor / generic mcpServers config:
{ "mcpServers": { "sfxmint": { "url": "https://sfxmint.com/mcp" } } }
To let the agent make sounds too, add your API key (only the generation tools use it):
claude mcp add --transport http sfxmint https://sfxmint.com/mcp --header "Authorization: Bearer $SFXMINT_API_KEY"
{ "mcpServers": { "sfxmint": { "url": "https://sfxmint.com/mcp", "headers": { "Authorization": "Bearer sfx_live_…" } } } }
Prefer a skill over a server? This repo also ships an Agent Skill that uses the plain HTTP API — no MCP server required:
npx skills add flreey/sfxmint-mcp -y
The agent chooses a role for one event, a set for several events, or structured search for a description or gap. These are task branches, not a required sequence. Library files, other permitted sources and local synthesis can be mixed. The skill includes React / Phaser / plain-HTML integration examples. Source: skills/sfxmint/SKILL.md.
No agent involved? The sfxmint CLI puts the same files in a project directly:
npx sfxmint add coin jump hit
It writes the audio plus a sounds.json manifest recording each file's source URL, SHA-256 and license, and a SOUNDS.md table linking each sound's page, and ships a small Web Audio runtime (createSoundboard) for first-gesture unlock, overlapping cues and a persisted mute. Source: flreey/sfxmint-npm.
Tutorials
Tools
| Tool | What it does |
|---|
get_role_sound | Ask by event — button-click, purchase-success, error, coin, notification. Returns a file-checked candidate plus alternates; plain-English aliases work |
get_sound_set | Candidates grouped by event and style (18 sets: ui-crisp, retro-game, checkout, platformer, ai-coding-tool, …) |
search_sounds | Free-text retrieval with explicit duration, loop and format requirements; hard constraints can produce no candidates. When nothing fits, user_message is one line for the user with a short link (https://sfxmint.com/s/…) to make the sound in AI Studio |
get_sound | Full metadata for one sound (incl. a short description, measured acoustics, license) |
generate_sound | Make a new sound from a description when the library has nothing suitable. Needs an API key and uses the account's 1 free try first, then credits or a plan on the account — a credit pack ($3.90 for 10 credits or $9.90 for 30, never expiring) or the unlimited plan ($19.90 for 30 days, unlimited sounds); one credit is one generation, one payment, nothing renews — each up to 4 CC0 WAV takes of up to 10 s, with an OGG link (ogg_url) when credits or a plan made it; one that makes no take does not count and gives its credit back. With variations_of (a finished take) it makes a variation set: 4 close takes plus a Godot .tres and a Unity script for random playback |
get_generation | Status of a generation and fresh download links for its takes |
change_sound | Change one take of a generation as it is, from the user's words or an action (too muddy, too close, make it loop): three versions, free, no generation used |
Game kits (projects of many sounds) closed on 2026-10-20: SFXMint makes one sound at a time. The kit tools are no longer listed, a call to one answers tool_retired with what to use instead, and projects made before still open and download in AI Studio.
The four library tools never need a key. When nothing fits, pass user_message on to the user and list the sound under a Missing heading in SOUNDS.md with its link. Generation needs an SFXMint API key in the Authorization header (free account); without one, generate_sound answers api_key_required with isError and the same short link. A key uses the account's 1 free try first (held to the browser the key was made in and the network it calls from), then credits or a plan: once it is used the answer is payment_required with buy_url, for the user to choose one (pricing); an agent never pays. Ask the user before making sounds, and let them listen to the takes. The same generation is available over REST as POST https://sfxmint.com/api/v1/orders (docs). get_job and POST /api/v1/generate stay retired; existing job records remain readable at GET /api/v1/jobs/{job_id}.
Requirements, processing and acceptance
Use REST structured search (response=structured&limit=3) or the matching MCP parameters to pass required duration, loop and format explicitly. Check returned requirements; do not substitute near_matches. When nothing matches every word, closest lists the nearest sounds that meet every hard requirement, each with missing_terms, and library_fit says how well the library covers the request. Scores and match types describe ranking, not listening confidence. Legacy loopable does not prove a seamless loop.
Sets accept format=wav|mp3: download_urls follows the requested format, legacy urls stays MP3, and missing_roles identifies incomplete sets. File decoding and hashes do not certify scene suitability.
Task-permitted trimming, filtering and level changes are allowed. Retain originals, source URLs, license, processing recipes and new file hashes; re-measure requirements and preview the result. Changed bytes and new scenes do not inherit the original acceptance. Generating (generate_sound, the website, local synthesis or an external service) remains a separate, task-authorized choice.
Use URLs returned by the live API, without inventing or rewriting them. See the API docs, Phaser guide, React guide, and download example.
Rate limits & fair use
- Search/metadata: no auth, generous limits
- Hotlinking is allowed and encouraged; files are served with long-lived immutable caching
Contact
[email protected] · https://sfxmint.com
