AI agents

Volito ships an MCP server. Connect it and your agent can find a graphic, read its props, see a preview as an image, check it at every aspect and render it, or write a new one.

Connect it

Claude Code:

claude mcp add volito -- npx -y volito mcp

Cursor, Windsurf, Codex and other MCP clients take the same command in their config:

{ "mcpServers": { "volito": { "command": "npx", "args": ["-y", "volito", "mcp"] } } }

The server runs on stdio and needs Node 20.10 or newer. Video renders need ffmpeg on PATH; previews, stills and checks do not.

The tools

Tool	What it does
`motion_list`	Find graphics by words (`search`) or `category`.
`motion_info`	An item's props and parts; `code: true` adds its source.
`motion_styles`	The brand styles.
`motion_moves`	Every move, ease, idle, hit and mood the `motion` settings take.
`motion_preview`	A contact sheet as an image: the agent sees how it moves and whether the text fits. About 0.3 s.
`motion_still`	One frame as an image, written to `out` too when given.
`motion_check`	Errors and warnings at every aspect: a throw, a frame that does not parse, a graphic that draws nothing, text off the frame, words never spoken, props that do not fit.
`motion_render`	Render to a file and return its path. `format` is `mov` by default; without `out` it writes to `motion-renders/` in the server's working folder.
`motion_api`	The program API and the motion kit, for writing new graphics.

The preview, still, check and render tools take an item or the agent's own code (with placement for code), plus props, style, colors, aspect, duration, words, beats, motion, mood, seed and speed.

The loop an agent runs

Find the item. motion_list with a few words: “name plate”, “countdown”, “coupon”.
Read its props. motion_info gives every prop's type, default and length limit, and the parts it can animate.
Fill it in and pick the look. Props from the brief, a style name or the creator's DESIGN.md, an aspect and a duration.
Look before rendering. motion_preview returns an image the agent can see: does the text fit, does it read on footage, does it land in time? motion_check catches what the eye misses.
Render. motion_render writes the file and returns its path; the agent tells the user where it is.

When no item fits, the agent writes one: motion_api prints the runtime, and code goes to the same preview, check and render tools. Write your own has the rules that keep a program renderable.

The agent skill

A ready-made skill ships in the package at skill/SKILL.md: the loop above, the packs, the flags and the rules, written for an agent. Copy it into your agent's skills folder; for Claude Code:

npm install volito
cp -r node_modules/volito/skill ~/.claude/skills/volito

Sandboxed by default

Agent-written code runs in a fresh vm context with no network, no require and no eval, and a frame that runs too long is stopped. Only strings and numbers cross into it. It keeps mistakes contained; it is not a security boundary against a hostile pack.