# MCP server

> Let Claude, your IDE, or any MCP client drive Decodx directly.

# MCP server

Decodx exposes an MCP server, so an AI agent can create projects, add scenes, set narration, render
and share — using real tools, not a chat box.

## What an agent can do

The server covers the same surface a person has in the web app: create and update projects, add
sources, add and reorder scenes, set narration, choose voices and avatars, apply templates and brand
settings, estimate and run renders, create shares, and read back jobs, videos and documents.

## Connecting a client

The server lives at `https://api.decodx.ai/mcp`. You sign in once through your browser (OAuth); the
client never sees your password.

- **Claude (web and desktop):** Customize → Connectors → **Add custom connector**, paste the URL, then
  **Connect**. On Team and Enterprise, an owner adds it under Organization settings → Connectors.
- **Claude Code:** `claude mcp add --transport http decodx https://api.decodx.ai/mcp`, then run `/mcp`
  to sign in. Or install the plugin, which also adds the decodx skill:
  `claude plugin marketplace add decodx-ai/claude-plugin` then `claude plugin install decodx@decodx`.
- **ChatGPT:** with developer mode on (Settings → Security and login), add a plugin with the same URL.
- **Any other MCP client:** point it at the URL; it discovers sign-in from the server's OAuth metadata.

The **AI assistants** page in the app (second in the sidebar, marked MCP) shows the same steps and
lets you review or revoke an agent's access.

## How an agent builds a video that renders first time

1. Create a project and add scenes. Every change replies with a list of anything still wrong.
2. `preview_scene` shows frames of a slide, title, terminal or motion scene, plus any layout problem —
   text overflowing its box or clipped by the edge — using the same check the render runs.
3. `validate_project` runs every pre-render check and returns all issues at once, each with a fix.
4. `estimate_duration` shows the length and the AI minutes a render will use.
5. `render_project` runs the same checks and only starts a job when they pass; `get_job` reports progress.
   Rendering spends AI minutes, so by default a person approves it first: the assistant gets a review
   link to share, and `get_render_request` tells it when the render was approved. Admins can let
   assistants render up to a daily limit, or always, in Settings → Workspace.

The [decodx-video skill](https://decodx.ai/skills/decodx-video/SKILL.md) teaches an agent this loop.

## Safe retries and concurrent edits

Every change accepts an optional `idempotency_key`: if a call times out, retry it with the same key and
it is not applied twice. Pass the project `version` (from `get_project` or the last change) as
`expected_version`, and a change made meanwhile by someone else is never overwritten — the call is
refused and the agent re-reads the project.

## Resources and prompts

Clients can attach read-only **resources** without a tool call: the authoring guide with every
pre-render check (`decodx://guide/authoring`), the validation rules, the slide templates, voices,
avatars, and any project's scenes (`decodx://projects/{project_id}/composition`).

**Prompts** are ready-made recipes in your client's prompt menu: *Product demo video*, *Turn a slide
deck into a video*, *Explainer from notes*, and *Recording to tutorial*.

## Authorising an agent

Agents authorise through OAuth: the agent gets its own grant that you can review and revoke,
scoped to your workspace, rather than a shared password.

For unattended automation — a script, a CI job, a backend service — use a
[service token](/docs/automate/rest-api) instead.

## Why this is worth using

An agent driving Decodx produces a project you can open, read, edit and re-render. The output isn't
a one-shot generation you either accept or throw away — it's an artifact that a person can take over
at any point.