Connect Vertov to your AI

The Vertov MCP server lets Claude, ChatGPT, Codex and other MCP clients read the videos you can see, as transcripts and briefs, and upload recordings to get a share link back. You sign in with your Vertov account and choose one workspace; the app acts as you there. No tool records your screen: an agent uploads video it already has. (Owners and admins can also have an agent record a narrated how-to of their own verified web app in a hosted browser with record_tutorial; that is not a recording of your screen.) For what this is for, see Vertov for founders and builders. A short written walkthrough of the Settings side is in how to connect an AI agent to your videos.

From a recording to your agent

These are the steps Vertov supports today, with what has been checked at each one. Your agent does the work; Vertov supplies what it needs to read and a place to send the result. For the overview, see Vertov for founders and builders.

  1. 1. Record the issue or the change you want

    Verified

    Record in the web recorder or the Chrome extension and say what is wrong or what should change. Stopping uploads the recording and transcribes it.

    A full recording, transcript and brief were checked on production with a test account on 4 November 2026. The extension's click capture was only tested with simulated clicks.

  2. 2. Copy a Markdown brief for your agent

    Verified

    In the video's Share menu, Copy for an AI agent puts a Markdown brief on your clipboard to paste into Claude, Cursor or any agent. Anyone who can open the video can copy it, and the same text is served at /api/prompt/<video id> and by the MCP tool get_video_prompt.

    The brief was generated and read back through the share menu, the API and the MCP tool on production, 4 November 2026. A private video's brief needs the same access as the video, so a script needs an open link or a signed-in session; MCP covers that case.

    Tools:get_video_prompt

  3. 3. The brief is built from the transcript

    Verified

    It holds the title, link and length, the summary, the chapters, the pages and buttons the extension captured when click capture was on, and the transcript with a [m:ss] stamp on every line. Each stamp maps to ?t=<seconds> on the video link, so the agent can point you to a moment.

    A spoken test clip was transcribed and its brief read back on production on 4 November 2026. A silent recording gets no transcript or summary, so an agent's silent Playwright video has only a title and link. Files over 25 MB sent through the API or copied from a URL are not sent for transcription.

  4. 4. Connect the agent through MCP or an API key

    Built, not fully tested

    Add https://vertov.app/mcp to Claude, ChatGPT or Codex and sign in, or give a script an API key from Settings, under Agents. Sign-in shows a consent screen (with a workspace choice if you have several) and can be undone in Settings, under Connected apps.

    API-key calls were exercised on production. The sign-in (OAuth) flow was tested locally and its discovery documents on production, but not yet inside the real Claude, ChatGPT or Codex apps, so those steps are unverified.

  5. 5. The agent reads the recording and transcript

    Verified

    It can list and search your videos, read a video's summary, chapters and transcript, fetch the brief, and list spaces. It sees only what your account can open: your own videos and anything in a space, and workspace owners and admins can open any video in the workspace by its link.

    Each read tool was called over production MCP with a test account on 4 November 2026. Listing without a search shows only your own videos that are not in a space or folder; searching also covers videos in spaces.

    Tools:list_videosget_videoget_video_promptlist_spaces

  6. 6. The agent writes and reads AI documents

    Built, not fully tested

    From a video's transcript an agent can write an SOP, bug report, meeting notes, changelog, handoff, action items, summary or message, and read one that already exists. Only the video's owner or a workspace owner or admin can; writing replaces an earlier document of the same kind.

    AI documents were generated and read on production on 4 November 2026, and get_ai_doc was among the read tools called over MCP. A generate_ai_doc call over MCP is not separately logged. Every account can write them during early access; they are planned for paid plans.

    Tools:generate_ai_docget_ai_doc

  7. 7. Your agent does the work

    Not built

    Vertov does not write code or run your agent. Your coding agent makes the change from the brief; whether it can finish a task depends on the agent and the task.

    No end-to-end run of an agent fixing something from a Vertov brief has been recorded, so we make no claim about results.

  8. 8. The agent uploads a recording of its work

    Verified

    An agent that already has a video file, for example from Playwright, a screencast or ffmpeg, sends it as base64 (the limit is about 15 MB, but the model writes every character, so keep it short) and gets a share link back. Larger files go through the API with a key. The upload uses your workspace's default sharing setting, so check it before a client link goes out.

    Uploading a video inline over MCP was tested on production on 4 November 2026 with valid, invalid-base64 and no-data calls. No tool starts a recording of your screen (start_tutorial, a pilot, only records Vertov's own screens on an owner-run runner): an agent needs its own recording source, and Playwright's recordings are silent.

    Tools:upload_video

  9. 9. Other ways the agent can answer

    Built, not fully tested

    It can copy a video from a public https URL into your library, and add a comment, with an optional timestamp, as you.

    Copy-from-URL has no logged production run and a failed download leaves the video marked failed. A comment was part of the 4 November 2026 production tests, but the log does not say it went through the MCP tool, so it is not counted as verified.

    Tools:create_video_from_urladd_comment

  10. 10. The whole loop, with a real agent

    Built, not fully tested

    You record a request, the agent reads the brief, does the work and returns a recording you watch and comment on.

    Each half has been exercised separately. A run of the whole loop with a real coding agent, recorded for this page, has not been made yet.

What an agent cannot do

  • No tool records your screen. An agent uploads a file it already has. Separately, owners and admins can have an agent record a narrated how-to of their own verified web app in a hosted browser (record_tutorial); that is not a recording of your screen.
  • No tool deletes a video or changes who can see it.
  • Inline uploads stop at about 15 MB; larger files use the API with a key (or a public URL).
  • A silent recording has no transcript or summary.
  • A connection acts as you in one workspace and sees only what you can open.

Server address

https://vertov.app/mcp

Streamable HTTP, stateless, JSON responses (a GET returns 405). Signing in with OAuth needs a full Vertov account, so a guest signs up first; the Free plan is enough. Scripts can use an API key instead.

Claude

In Claude, open Customize, then Connectors, and add a custom connector with the address above (the directory listing is not live yet). Choose Connect, sign in to Vertov and select Allow.

In Claude Code:

claude mcp add --transport http vertov https://vertov.app/mcp

Then run /mcp in Claude Code and choose Authenticate.

ChatGPT and Codex

In ChatGPT, add the address above as a custom MCP connection, then sign in; the apps directory listing is not live yet. In the Codex CLI, add the server, then sign in when Codex asks:

codex mcp add vertov --url https://vertov.app/mcp

The sign-in flow has been tested locally. It has not yet been run inside the real Claude, ChatGPT or Codex apps, so if a step differs from the above, tell us.

With an API key instead

Scripts and clients without OAuth can send an API key from Settings, under Agents:

claude mcp add --transport http vertov https://vertov.app/mcp \
  --header "Authorization: Bearer rk_your_key"

A key acts as you in one workspace. Keep it out of screenshots, recordings and prompts; revoke it in the same place.

Try it without a client

Any HTTP client can speak to the server. Start with initialize, then list the tools, then call one:

curl -sS https://vertov.app/mcp \
  -H 'Authorization: Bearer rk_your_key' -H 'Content-Type: application/json' \
  -d '{"jsonrpc":"2.0","id":1,"method":"initialize","params":{"protocolVersion":"2025-06-18","capabilities":{},"clientInfo":{"name":"curl","version":"1"}}}'
# same headers, body: {"jsonrpc":"2.0","id":2,"method":"tools/list"}
curl -sS https://vertov.app/mcp \
  -H 'Authorization: Bearer rk_your_key' -H 'Content-Type: application/json' \
  -d '{"jsonrpc":"2.0","id":2,"method":"tools/call","params":{"name":"get_video_prompt","arguments":{"video":"ab12cd34"}}}'

A call returns its result as text in result.content; a failed call sets isError with the reason.

Tools

ToolEffectWhat it does
list_videosReadList your recent videos, or search titles, summaries and transcripts.
get_videoReadA video's title, status, summary, chapters and timestamped transcript.
get_video_promptReadA video as a Markdown brief of what was shown and said, with any captured clicks.
get_ai_docReadA document already written from a video (SOP, bug report, meeting notes and more). Owner or manager only.
list_spacesReadThe shared spaces in the workspace.
upload_videoWriteUpload an MP4 or WebM (up to about 15 MB inline) and get its share link.
create_video_from_urlWriteCopy a video from a public https URL into your library.
add_commentWriteComment on a video as you, optionally at a timestamp.
generate_ai_docWrite, replacesWrite a document from a video's transcript, replacing an earlier one of the same kind. Paid plans.
inspect_app_pageReadA screenshot and the buttons, links and fields of a page of your own app, signed in with its demo login. Owners and admins.
record_tutorialWriteRecord a narrated how-to video of your own app from steps the agent writes; it lands in your library. Owners and admins.
rerecord_tutorialWriteRecord an earlier tutorial's steps again, e.g. after a release.
list_tutorial_recipesReadTutorials a runner on the owner's computer can record. Owners and admins.
start_tutorialWriteQueue a narrated tutorial recording for that runner. Owners and admins.
get_tutorial_jobReadA tutorial job's progress and, when done, its video link; a failed recording includes the error and a screenshot.
cancel_tutorial_jobWrite, finalCancel a queued or running tutorial job.

There are no tools that delete videos or change who can see them.

Copy for an AI agent

No connection is needed to hand an agent a brief. In a video's Share menu, choose Copy for an AI agent and paste the Markdown into Claude, Cursor or any agent. It holds the title, link and length, a summary, chapters, the pages and buttons the Chrome extension captured (when click capture was on) and the transcript with a timestamp on every line. A timestamp [1:05] matches ?t=65 on the video link.

The same text is at /api/prompt/<video id> for anyone who can open the video. A private video needs the same access as the video itself, so for those use get_video_prompt through a connection. A silent recording has no transcript, so its brief has only the title and link.

Sending a recording back

An agent needs its own recording source: Playwright's recordVideo, a screencast or ffmpeg. Vertov turns the finished file into a share link; it does not record for the agent. New videos use your workspace's default sharing setting, so check it before sending a link to anyone outside the team. The model has to write every character of the base64 into the call, so inline upload suits short, small clips; for anything longer have the agent run the curl command with a key. Files over 25 MB uploaded this way are not transcribed, and Playwright's recordings have no audio, so expect no transcript.

# Inline, from an MCP client or script: base64 of the file, up to about 15 MB
{"name":"upload_video","arguments":{"title":"Checkout fix, verified","content_type":"video/webm","duration":42,"data_base64":"<base64 of the file>"}}

# Larger files: the API, with a key
curl -sS -X POST 'https://vertov.app/api/v1/videos?title=Checkout%20fix' \
  -H 'Authorization: Bearer rk_your_key' -H 'Content-Type: video/webm' \
  --data-binary @videos/run.webm

Examples

  • “Read vertov.app/ab12cd34 and fix the bug it shows.”
  • “Find the video where we discussed the pricing page and summarize the decisions.”
  • “Turn my onboarding walkthrough into a step-by-step SOP.”
  • “Upload this screen recording and give me the link.”

Data and disconnecting

The connection can reach only the workspace you chose, and only videos you could already open there. Vertov doesn't receive your conversation; it gets the tool calls the app makes. Disconnect any time in Settings, under Agents, under Connected apps. The privacy policy and terms apply.

Help

Email support@vertov.app or see the support page. For plain REST calls and webhooks, see the API docs.

Commands and recent recordings
Actions
Go to
↑ ↓ Select↵ OpenEsc Close