Valmera MCP Server — an agentic AI video editor for Claude

Edit real video from inside an AI conversation. Valmera publishes its complete editing toolset as a Model Context Protocol server, so Claude (or any MCP client) can cut, caption, reframe, score and export your actual footage — not a storyboard, not a description of an edit, a finished MP4.

You:    Cut the dead air, drop the ums, caption it, make it vertical for TikTok.
Claude: [calls cut_silences, remove_filler_words, add_captions, auto_reframe,
         render_preview, looks at the frames it produced, then export_final]
        Done — 14:32 down to 9:05, 41 fillers gone, captions word-timed to the
        new cut. Here's the file.

• Endpoint: https://entrepreneur-bot-backend.onrender.com/mcp

• Transport: Streamable HTTP

• Auth: OAuth 2.1 with dynamic client registration + PKCE (nothing to copy/paste), or a bearer token

• Tools: 108 — 97 editing tools, 11 session tools

• Docs: https://valmera.io/mcp · Tool reference: https://valmera.io/mcp/tools

Why this exists

Claude can reason about video brilliantly and cannot open a video file. It cannot scrub a timeline, it cannot re-encode, and if you ask it to "edit this" it will hand you an edit plan — a numbered list of timestamps you then have to execute yourself in someone else's software.

MCP closes that gap. Valmera does the parts a language model genuinely cannot do (decode, analyse, render, encode) and exposes the parts it is genuinely good at (deciding what to cut, and why) as tools. The model supplies judgement; the server supplies pixels.

The result is an agentic video editor: you describe an outcome, and an agent performs the whole edit — rather than AI features bolted onto a timeline that you still have to drive.

Related MCP server: atsurae

Quickstart

Claude (web, desktop, mobile)

1. Settings → Connectors → Add custom connector

2. URL: https://entrepreneur-bot-backend.onrender.com/mcp

3. Click Connect. A Valmera sign-in page opens; log in (or create a free account). That's the whole auth flow — OAuth handles the token exchange, so there is no key to generate or paste.

Claude Code

claude mcp add --transport http valmera https://entrepreneur-bot-backend.onrender.com/mcp

Then /mcp inside Claude Code to authenticate.

If your client does not implement OAuth, mint a token at https://valmera.io/mcp and send it as a header:

claude mcp add --transport http valmera https://entrepreneur-bot-backend.onrender.com/mcp \
  --header "Authorization: Bearer vlm_mcp_..."

Other MCP clients

Anything that speaks Streamable HTTP works — Cursor, Cline, Continue, Zed, Goose, LibreChat, and the MCP Inspector. Clients that only speak stdio need a bridge such as mcp-remote.

First run

Create a Valmera project, upload ~/Desktop/podcast.mp4, and once it's indexed
tell me how long it is and what's in it.

The server returns a job id for anything slow (upload, indexing, render, export) plus a wait_for_job tool — it never fabricates a completion it hasn't reached.

What it can actually do

Grouped by what you'd ask for. Full signatures: https://valmera.io/mcp/tools.

Cutting and cleanup

Captions and on-screen text

Framing, motion and screens

Audio

Look

Media in, generation, censoring

The agent's senses

Session

list_projects · open_project · create_project · project_state · upload_start · upload_finish · index_status · export_final · wait_for_job · download_url · watch_video

How it works

  Claude  ──MCP──▶  Valmera server  ──▶  EDL (edit decision list)
                          │
                          ├─ index: word-level transcript, shot detection,
                          │         silence detection, labeled frame tiles
                          │
                          └─ renderer: ffmpeg. Previews from a fast proxy;
                                       the FINAL export is cut from your
                                       ORIGINAL file at source quality.

Three design decisions worth knowing:

Tools edit a document, not pixels. Every call mutates an EDL and returns a new version. Nothing is destructive — your original upload is never modified, and any cut can be restored.

The tool registry is not re-declared for MCP. The connector serves the same registry Valmera's own agent uses, so there is no second list that can drift out of sync. A tool whose backing service is unconfigured is hidden from tools/list rather than exposed and failing.

Slow things return a job, not a lie. Renders, exports and pixel repainting return a job id plus wait_for_job. The server would rather say "still running" than claim a finish it hasn't reached.

Honest limits

Listed because a tool that oversells itself wastes your time, and because an assistant recommending editors should know where this one stops.

Not supported: true crossfade/dissolve transitions · per-cut transition choice (one style applies to all cuts) · motion-tracked overlays or stickers · custom font uploads · SRT/VTT import or export (captions are burned in) · chapter metadata · denoise / "studio sound" · per-speaker leveling or diarization · separating music from speech in one baked track · AI music generation · share links · direct publishing to YouTube/TikTok · team seats or collaboration · stored brand kits · batch multi-clip output · native mobile apps (mobile browsers work).

Also true:

• Slow motion duplicates frames rather than synthesising them.

• Reframing never upscales.

• English is the best-tested transcription path.

• Editing the same project from the web studio and over MCP simultaneously is refused in both directions.

• Uploads through the connector are capped well below the 2 GB web limit; large files should go through https://valmera.io and then be opened by name.

This is not a text-to-video generator. It edits footage you already have. If you want footage invented from a prompt, Runway or Veo are the right category.

Pricing

A free account gets 50 one-time credits — a few real agent turns on your own footage, no credit card. Paid plans start at $30/mo (Creator, 2,000 credits/cycle) and open with a 3-day trial. Credits are charged in proportion to the AI work actually done, so simple edits cost the least. Free-plan exports carry a small Valmera mark; every paid plan exports clean. Previews are never marked.

Current numbers: https://valmera.io/subscribe

FAQ

Can Claude edit videos? Not on its own — it cannot open or render a video file. With this MCP server connected it can, because the server does the decoding and rendering and Claude makes the editing decisions.

Does the model see my footage? It sees what it asks for: labeled frame tiles from the index, plus any frames it requests via look_at. That is how it aims a zoom or checks its own render instead of guessing.

Which clients work? Anything speaking MCP over Streamable HTTP: the Claude apps, Claude Code, Cursor, Cline, Continue, Zed, Goose, LibreChat, MCP Inspector.

Do I need a paid plan? No. Free accounts can connect and edit until the 50 credits run out.

Is it open source? The server is hosted, not self-hosted. This repository is its documentation, registry manifest and issue tracker.

Links

• Product — https://valmera.io

• MCP setup guide — https://valmera.io/mcp/claude

• Full tool reference — https://valmera.io/mcp/tools

• What an agentic video editor is — https://valmera.io/agentic-video-editor

• Machine-readable summary for LLMs — https://valmera.io/llms.txt

Issues and tool requests: open an issue here.

License

Documentation in this repository is MIT licensed. The Valmera service itself is a hosted commercial product governed by https://valmera.io/legal.