mcp

<p align="center">
  <a href="https://rendobar.com">
    <picture>
      <source media="(prefers-color-scheme: dark)" srcset="https://cdn.rendobar.com/assets/brand/logo-mark.svg">
      <img alt="Rendobar" src="https://cdn.rendobar.com/assets/brand/logo-mark-black.svg" width="80">
    </picture>
  </a>
</p>

<h1 align="center">@rendobar/mcp</h1>

<p align="center">
  <strong>Serverless media processing for AI agents.</strong><br>
  The official Model Context Protocol server for Rendobar.
</p>

<p align="center">
  <a href="https://rendobar.com/docs/mcp-server">Docs</a> &nbsp;·&nbsp;
  <a href="https://www.npmjs.com/package/@rendobar/mcp">npm</a> &nbsp;·&nbsp;
  <a href="https://glama.ai/mcp/servers/kwdj3f0u3z">Glama</a> &nbsp;·&nbsp;
  <a href="https://discord.gg/kAGqjBzx8N">Discord</a>
</p>

<p align="center">
  <a href="https://www.npmjs.com/package/@rendobar/mcp"><img src="https://img.shields.io/npm/v/@rendobar/mcp?style=flat-square&color=059669&label=npm" alt="npm version"></a>
  <a href="https://www.npmjs.com/package/@rendobar/mcp"><img src="https://img.shields.io/npm/dm/@rendobar/mcp?style=flat-square&color=059669" alt="npm downloads"></a>
  <img src="https://img.shields.io/npm/l/@rendobar/mcp?style=flat-square&color=059669" alt="MIT license">
  <img src="https://img.shields.io/node/v/@rendobar/mcp?style=flat-square&color=059669" alt="Node version">
</p>

`@rendobar/mcp` is the official Model Context Protocol server for [Rendobar](https://rendobar.com), a serverless media processing API. The server runs locally over stdio and reads files straight from your disk, so an AI agent can take a file off your machine, process it on Rendobar's infrastructure, and hand back a hosted URL.

Rendobar covers both sides of media work.

**Transform what you have.** Run any FFmpeg command against video, audio or images the way you would write it locally. Inspect a file and get a normalized summary plus the full ffprobe report. Compose video from a declarative JSON timeline. Compress to a target size or quality, where the encoder searches candidate encodes and returns the smallest file that clears the bar. Burn in subtitles from SRT, VTT or ASS, or let it transcribe when none is given.

**Generate what you do not.** Create an image from a text prompt on hosted open-weight diffusion models. Edit up to four reference images from a written instruction, no masks and no coordinates. Upscale on a one-step diffusion restoration model that reconstructs detail rather than only sharpening. The same model-backed layer drives the transcription and keyword highlighting behind animated captions, so this is not an image-only capability.

The job list grows over time, so this README names families rather than types. `list_job_types` reads the current set live from the registry on every call.

Published to npm as `@rendobar/mcp` and to the [official MCP Registry](https://registry.modelcontextprotocol.io) as `com.rendobar/mcp`.

## Without it

> **You:** Mute the first 3 seconds of `intro.mp4`.

The agent tells you to install FFmpeg. Then you go looking for how to gate a
filter on a timestamp, land on `volume=enable='lt(t,3)'`, and lose another few
minutes to quote escaping in your shell. Nobody remembers that syntax, which is
the problem.

## With it

> **You:** Mute the first 3 seconds of `intro.mp4`.

```jsonc
upload_file  { "path": "~/clips/intro.mp4" }
// → { "downloadUrl": "https://cdn.rendobar.com/u/abc123/intro.mp4", "sizeBytes": 4821004 }

submit_job   { "type": "ffmpeg",
               "inputs": { "intro.mp4": "https://cdn.rendobar.com/u/abc123/intro.mp4" },
               "params": { "command": "-i intro.mp4 -af \"volume=enable='lt(t,3)':volume=0\" -c:v copy out.mp4" } }
// → { "jobId": "job_9f2a", "status": "waiting" }

get_job      { "jobId": "job_9f2a", "wait": true }
// → complete · $0.01 · https://cdn.rendobar.com/o/job_9f2a/out.mp4
```

The agent writes the filter. Rendobar runs it. Nothing gets installed on your
machine, and `-c:v copy` means the video stream is never re-encoded.

## Two more things to ask for

**Hit a size budget.**

> **You:** Get `demo.mov` under 25 MB so I can email it.

```jsonc
upload_file  { "path": "~/recordings/demo.mov" }
// → { "downloadUrl": "https://cdn.rendobar.com/u/7c1e/demo.mov", "sizeBytes": 251658240 }

submit_job   { "type": "compress.target",
               "inputs": { "source": "https://cdn.rendobar.com/u/7c1e/demo.mov" },
               "params": { "for": "web", "target": { "maxSize": "25MB" } } }
// → { "jobId": "job_4b8d", "status": "waiting" }

get_job      { "jobId": "job_4b8d", "wait": true }
// → complete · https://cdn.rendobar.com/o/job_4b8d/out.mp4 · 23.8 MB
```

You give it the ceiling, not a bitrate. The encoder searches candidate encodes
and returns the smallest file that still clears the quality bar, so you are not
guessing at CRF values to land under a mail server's limit.

**Generate an image.**

> **You:** Make a 1920x1080 title card for a video about deep sea diving.

```jsonc
submit_job   { "type": "image.generate",
               "inputs": {},
               "params": { "model": "standard",
                           "prompt": "Title card for a deep sea diving documentary. Shafts of light through deep blue water, small diver silhouette, empty space across the upper third for a title.",
                           "width": 1920, "height": 1080 } }
// → { "jobId": "job_2fa7", "status": "waiting" }

get_job      { "jobId": "job_2fa7", "wait": true }
// → complete · https://cdn.rendobar.com/o/job_2fa7/out.png
```

`inputs` is empty because nothing is being transformed. Ask for a tier
(`economy`, `standard`, `premium`) and the platform picks the model, or pin an
exact model id to reach its own controls. Requested dimensions are snapped to
what the chosen model can actually render.

## The rest of the surface

Six more tools, and the prompts that reach them.

> **You:** What can Rendobar actually do?

```jsonc
list_job_types {}
// → { "jobTypes": [ { "type": "compose", "tag": "Compose",
//                     "summary": "Render a video from a declarative JSON timeline",
//                     "acceptsMedia": ["video", "image", "audio"] }, ... ],
//     "guidance": "..." }
```

Read live from the job registry on every call, which is why nothing in this
README enumerates job types. A new one appears here without a release.

> **You:** How much credit is left?

```jsonc
get_account {}
// → { "balance": "$4.86", "balanceUsd": 4.86, "plan": "free", "isPro": false,
//     "limits": { "concurrentJobs": 1, "maxFileSize": "500 MB", "jobTimeoutMin": 60 } }
```

Worth a call before submitting something expensive.

> **You:** What did I run this morning?

```jsonc
list_jobs { "status": "complete", "limit": 5 }
// → { "jobs": [ { "id": "job_9f2a", "type": "ffmpeg", "status": "complete",
//                 "createdAt": "2026-08-04T09:12:00Z", "cost": "$0.01",
//                 "output": { "url": "https://cdn.rendobar.com/o/job_9f2a/out.mp4" } } ] }
```

The compact row is enough to find a result you lost. Call `get_job` when you
need the full output.

> **You:** Stop that one, I picked the wrong file.

```jsonc
cancel_job { "jobId": "job_9f2a" }
// → { "id": "job_9f2a", "status": "cancelled" }
```

Works on `waiting`, `dispatched` and `running` jobs. A running job's upstream
execution is stopped too, so you are not billed for work you cancelled.

> **You:** Where can this land once it's done?

```jsonc
list_storage {}
// → { "storage": [ { "id": "prod-media", "provider": "s3", "bucket": "acme-prod-media",
//                     "access": "deliver", "defaultDestination": true, "pending": false } ],
//     "note": "..." }
```

The buckets connected on the Storage page, without the endpoint or region an
agent has no use for. Use the id in `destinations`, or as
`storage://<id>/<path>` in any input.

> **You:** What's already in the archive bucket?

```jsonc
list_storage_files { "storageId": "raw-archive" }
// → { "folders": [ { "prefix": "2026/", "uri": "storage://raw-archive/2026/" } ],
//     "files": [ { "key": "notes.txt", "size": 800, "lastModified": "2026-09-01T00:00:00.000Z",
//                  "uri": "storage://raw-archive/notes.txt" } ],
//     "cursor": null }
```

Every entry carries its own `storage://` uri, ready to hand straight to
`submit_job`. Pass a folder's prefix to go one level deeper.

## Install

Rendobar has two MCP servers. Pick by whether the agent needs your filesystem.

| | `@rendobar/mcp` (this package) | Hosted (`api.rendobar.com/mcp`) |
|---|---|---|
| Transport | stdio, spawned by your client | Streamable HTTP |
| Reads local files | Yes. That is the reason it exists | No. The server has no disk |
| Auth | API key | OAuth in the browser, or a Bearer key |
| Best for | Claude Desktop, Cursor, Cline, Zed | claude.ai, ChatGPT, hosted gateways |

**Hosted, no API key**, one command:

```bash
claude mcp add --transport http rendobar https://api.rendobar.com/mcp
```

**Local**, for filesystem access. Get a key at [app.rendobar.com](https://app.rendobar.com) → Settings → API Keys, then:

```bash
claude mcp add rendobar -s user --env RENDOBAR_API_KEY=rb_... -- npx -y @rendobar/mcp
```

Already ran `rb login` with the Rendobar CLI? Drop `--env`. The server finds the credentials file.

<details>
<summary><strong>Claude Desktop, Cursor, Cline, Windsurf</strong></summary>

Same block for all four. Claude Desktop: `~/Library/Application Support/Claude/claude_desktop_config.json` (macOS) or `%APPDATA%\Claude\claude_desktop_config.json` (Windows). Cursor: `~/.cursor/mcp.json` on every OS. Windsurf: `~/.codeium/windsurf/mcp_config.json` on every OS. Cline: MCP panel → Configure.

On Linux, use Cursor, Windsurf, Cline, Zed, VS Code or Continue. Claude Desktop has no Linux build, so it is the one client on this list you cannot use there. The server itself runs fine on Linux.

```json
{
  "mcpServers": {
    "rendobar": {
      "command": "npx",
      "args": ["-y", "@rendob