DemoBitesManage DemoBites from your agent
DemoBites is not a video tool, it is the platform that runs your whole demo and update flow, and the management MCP is the proof. Your agent can draft releases, assign bites, manage languages, and publish, with you approving anything public.
Your customers’ agents read your centers. Your agent runs them.
Every published center already serves its own read side for AI agents. This is the write side.
Quickstart
One command registers the management MCP with your agent, and runs the sign in first if you are not connected yet:
npx demobite mcpIf you already ran npx demobite, you are done, the front door sets up the recorder skill and the MCP together. You never have to choose between them. Inside the app, the same setup lives under MCP in your account menu.
From claude.ai or ChatGPT
Browser assistants need no terminal and no key. In the connector settings, add an MCP server pointing at one URL:
A DemoBites approval page opens in your browser, you pick a workspace, click Approve, and the assistant is connected. Under the hood it is OAuth 2.1 with PKCE and dynamic client registration, which is exactly why the URL is all you need.
Any MCP client
The server speaks Streamable HTTP, so any MCP client that can send a URL and a header can connect, Claude Code and Cursor among them. Register it manually with:
No key yet? Connect without one and the connect_demobites tool walks your agent through linking your account, it returns a code you approve in the browser, and the key comes back to the agent.
Codex, Claude Desktop, Gemini and any other client that speaks Streamable HTTP should connect the same way; we have hands on verified Claude Code and Cursor. Done with a machine? npx demobite logoutrevokes its key server side. A hosted connector stops when you remove it in the assistant’s settings.
What your agent can do
One hundred and seven tools across nine groups, your whole workspace by tool call. Twenty seven of them touch something public, and those pause for the approve moment below.
Connect and status
5How an agent introduces itself, links your account, and checks where things stand.
Bites
13List, rename, tag, organize, duplicate, share and retire your demo videos.
Timeline
12The studio timeline by tool call, narration beats, cuts and camera.
Languages and narration
16Add languages, manage narration and voices, stage new editions.
Spotlight
9Cut, configure and fetch the finished renders of your spotlights.
Update Center
20Releases end to end, create, edit, assign bites, publish, all of it.
Explore Center
7The evergreen lineup, slots, settings and publishing.
Analytics and subscribers
5Read your analytics summary, manage destinations and subscribers.
Workspace and account
20Brand kit, domains, tags, folders, workspace settings.
The full list, every tool with its inputs and approval markers, lives on the MCP tool reference, generated from the live server.
The server documents itself: a tools/list call returns every tool with its full input schema, so what your agent sees is always current.
The approve moment
Your agent drafts freely, but nothing reaches a public page without you. Any publishing tool answers its first call with a human readable preview and a confirmation token instead of acting. Your agent shows you the preview, you say yes, and only the second call with that token executes. Tokens are single use, bound to that exact action and its arguments, and expire after ten minutes.
Failures are built for agents too: every error comes back as a structured object carrying the error name, a human message, and the fix, expired or reused confirmation tokens included, so a good agent recovers without guessing.
On the record
Every management call is audit logged, who connected, what was called, what changed. Agency without accountability is not a feature.
Recording vs managing
Two different jobs, two different doors. For recording, use the Agentic Recorder skill whenever your agent has your repo: an agent that knows the code films real selectors and real flows, with fewer wrong turns. The cloud recorder covers everyone else. For managing what you publish, the MCP is the door for any agent, repo or no repo.
Plans
Keys and tokens carry two scopes. The record scope connects an agent and covers the connect and status tools. The manage scope covers everything inside a workspace, reading included, and it takes both the scope and a workspace on Grow.
- Connecting an agent is open on every plan.
- Everything your agent does inside a workspace, reads included, requires Grow.
- Recording stays available on every plan, the gate is on managing, not creating.
- add_language follows your plan's localization feature, Global and up.
Straight answers
Which agents does this work with?
Coding agents through npx demobite, Claude Code and Cursor among them, and browser assistants like claude.ai and ChatGPT through the connector URL alone. Any MCP client that speaks Streamable HTTP can join.
Can the agent publish without me?
No. Every tool that touches a public page returns a preview and waits for your confirmation, and the token it needs is single use and expires.
Is what my agent does recorded?
Yes. Every management call is logged server side, who connected, what was called, what changed, denials included.
New here? Set up recording first, the same front door installs both.
Continue to the Agentic Recorder