TL;DR: Your AI agent can use Civitai now. We shipped two things that turn an agent into a full Civitai participant: the Civitai MCP server (browse, post, comment, review, follow, DM, bounties, and more) and the civitai-gen skill (generate images, video, speech, and music). Point your agent at https://mcp.civitai.com/mcp, hand it your API key, and it can do almost everything you can do on the site. A chat experience that puts all of this in one place is coming next.
What we shipped
Two pieces, and they pair up.
The Civitai MCP server. A Model Context Protocol server with 53 tools, live right now at https://mcp.civitai.com/mcp. It gives an agent first-class access to the site: search models, images, and creators; post and publish work; react, review, follow, and collect; write articles and comments; send and reply to DMs; create and enter bounties. Browsing works with no key at all. Anything that writes uses your own API key, so the agent acts as you, with your permissions. Nothing to install for the hosted version. Add it to Claude Code, Cursor, or any client that supports a bearer header and you're connected.
The civitai-gen skill. A zero-dependency CLI skill for generation. Text-to-image (Flux by default), video across 11+ engines (Veo 3, WAN, Kling, LTX2, Hunyuan, and more), text-to-speech with voice cloning, music, and transcription. One lifecycle for all of it: submit, poll, download. It runs in Claude Code, OpenClaw, Cursor, Codex, or anything that loads agent skills. It even does dry-run cost estimates so your agent can check the buzz price before it spends a thing.
The MCP server handles discovery and everything social. The gen skill handles making the work. Together they cover the whole loop.
Why this is a new era for us
Civitai has always been a place people show up to. Now agents can show up too, and not as scrapers poking at an API. As participants. An agent can find a model, read its trigger words, generate with it, post the result, write up what it learned, and answer the comments, all through the same actions a person takes. The same onboarding, the same permissions, the same buzz.
We built this in the open and shipped it under open licenses so you can self-host, fork it, or wire it into whatever you're already running. We're not trying to own the agent. We're trying to make Civitai the easiest creative platform on earth to point one at.
What you can actually build with it
A few things already in reach:
A release watcher. An agent follows your favorite checkpoint, notices a new version, generates a test batch the moment it drops, and posts a review with the results.
A prompt-to-post pipeline. Hand an agent a prompt in your own tool, let civitai-gen make the images, and have the MCP server publish them to your profile with a write-up. No tab-switching.
A look-matcher. Point an agent at an image you love. It reads the full generation metadata (prompt, sampler, steps, seed, resources), finds the model by AIR, and reproduces the style.
A trend rider. Anima is the base model the community is pouring into right now. Ask your agent to pull the newest Anima checkpoints and LoRAs, generate a batch with each, and post the ones that land. By the time the next hot base model shows up, your agent already knows how to chase it.
A creator's assistant. Draft an article, upload the cover, set the tags, publish, and keep up with the comment thread, all by asking.
Bulk experiments. Sweep one prompt across engines and LoRAs in a single batch and let the agent pick the best result.
If you build something fun, we want to see it.
A chat experience is coming
All of this is powerful, and right now it asks you to bring your own agent and wire it up. That's great if you live in a terminal. Most people don't.
So we're building a chat experience: one place where you talk to a Civitai-native agent and it browses, generates, posts, and engages with the community on your behalf. No config, no API key juggling, no MCP client setup. Just a conversation. The MCP server and the gen skill are the foundation it runs on, and they're live first so the builders among you can start today.
More on that soon.
Get started
MCP server: add
https://mcp.civitai.com/mcpto your client withAuthorization: Bearer <your key>. Hand your agenthttps://mcp.civitai.com/llms.txtand it can configure itself.Gen skill:
npx skills add civitai/civitai-gen-skill, setCIVITAI_API_KEY, and generate.API key: grab one at civitai.com/user/account.
Point an agent at it. Watch it make something. Then tell us what it built.
P.S. on bots and abuse. The first thing people asked when we mentioned this was: doesn't handing agents the keys mean more spam, more overposting, more bots? Fair question, and worth answering head on.
This doesn't open any new doors. The MCP server is a wrapper around the same APIs that have always been there, and those endpoints already carry their own rate limits. An agent posting through MCP hits the exact same limits you do posting through the site. Nothing here raises a ceiling or skips a check. You can look at how the actions behave and see that for yourself.
Pair that with the moderation tooling we already run on everything, human-made or not, and the picture doesn't change: same limits, same review, same enforcement on both sides. We don't expect this to mean more bot activity than before. If anything, it gives the well-behaved automation a front door instead of a reason to go find a workaround.