MCP server

The Finetuning MCP server lets AI agents work with your account directly: write a song, make a sound effect, check a playlist, fetch a certificate. It speaks the Model Context Protocol, so any MCP-capable client can connect.

https://mcp.finetuning.ai/mcp

Want the overview? See finetuning.ai/mcp.

Who can use it

API and agent access are included with Plus, Pro and Lifetime Tier 2 and above. That means a Plus or Pro subscription that is active (not canceled), or a Lifetime plan at Tier 2 or higher. It is the same rule as the API.

How authorization works

There is no API key to copy and paste.

1. Add the server URL in your agent. See Connect your agent.
2. The agent opens a Finetuning consent page in your browser. Log in once if you are not already.
3. Review what the agent is asking for and approve it.

The agent asks for four scopes: account, generate, library and certificates. Authorization uses OAuth 2.1 with PKCE, and agents register themselves automatically, so there is nothing to set up in advance.

Revoking access

Approving creates an entry in Profile → API, where the agent shows up as a key named MCP: <agent name>. Revoke it there at any time. The agent's next request is refused and it will ask you to reconnect.

What agents can do

• Check your account and credit balance.
• Write songs and instrumentals, and make sound effects.
• Wait for a generation to finish, then list and fetch your tracks and sound effects.
• Delete generations.
• Manage playlists: list them, add, remove and move tracks.
• Download the certificate for a track.
• List the curated style templates and read the song config.

See the tool reference for all 18 tools.

Reimagine, Remove vocals and uploads are not available through MCP. Those need your explicit rights confirmation, which only you can give, so use the web app or the API.

Costs

Agents spend what you would spend yourself:

• Tracks and instrumentals cost 1 generation each.
• Sound effects use the shared daily sound-effect cap, not credits.
• Every paid tool previews its cost, and honours a max_cost argument. If the cost is higher than max_cost, the call is refused before anything is created or charged.

Rate limits

The same as the API: 60 requests per minute per user, and 10 per minute for each kind of creation. See Rate limits.