Skip to main content
Glama
README.md
![An ad factory: one machine, a conveyor of UGC video ads, fanning out to every feed](docs/ad-factory.jpg)

# inReels MCP

**Your agent, with a video studio attached.** Paste one line into Claude, then
ask for an ad. You get back a finished MP4 — UGC video, product demo, faceless
short, static creative or voiceover.

```bash
claude mcp add --transport http --scope user inreels https://www.inreels.ai/api/mcp
```

Run `/mcp`, pick `inreels`, sign in in the browser. No API key to copy, no SDK
to install. Then just ask:

> **"Make me a 15-second UGC ad for this product: `<link or image>`"**

A new account starts with 15 free credits — about five clips — and there is no
subscription to begin.

---

## Three steps, about a minute

**1 — Connect**

| Client | |
|---|---|
| **Claude Code** | `claude mcp add --transport http --scope user inreels https://www.inreels.ai/api/mcp` |
| **Codex** | `codex mcp add inreels --url https://www.inreels.ai/api/mcp` then `codex mcp login inreels` |
| **Cursor** | in `~/.cursor/mcp.json`: `{ "mcpServers": { "inreels": { "url": "https://www.inreels.ai/api/mcp" } } }` |
| **Claude app · ChatGPT · Muse** | add a custom connector with the URL above |
| **Anything else** | same URL, plus OAuth or `Authorization: Bearer <key>` |

**2 — Sign in** from the browser when your agent asks. That's the whole setup.

**3 — Ask.** In plain English:

> "Make me a 15-second UGC ad for this product: `<link>`"
>
> "Three static ad variations for my cold-brew cans, 1:1 for Instagram"
>
> "A 60-second faceless video about the history of coffee, for Shorts"
>
> "Same ad again, but try four different hooks"

Your agent reads the prompt, picks the model, prices it, and comes back with
the file.

---

## What it costs

**1 credit = $0.10.** Pay for what you render; no subscription required.

| | Credits | |
|---|---:|---:|
| Image — FLUX schnell | 1 | **$0.10** |
| Voiceover — ElevenLabs, per take | 1 | **$0.10** |
| Video — Seedance Mini, 5s | 2–3 | **$0.20–0.30** |
| Video — MiniMax H3, 5s | 5 | **$0.50** |
| Image — Nano Banana | 4 | **$0.40** |
| Video — Kling 3.0 Standard, 5s | 11 | **$1.10** |
| Video — Veo 3.1, 4s | 17 | **$1.70** |

For comparison, one video from a UGC creator runs $200–500 and takes a week.

`quote` returns the exact price **before** anything is spent, and a failed run
is refunded in full. Read-only tools — browsing models, voices, actors, run
history — are free.

---

## No lock-in

This matters more than the feature list, so it is spelled out.

- **The server in this repo is MIT.** Fork it, read it, run it yourself.
- **Plain REST, same key.** Every tool is also an HTTP endpoint. No SDK, no
  proprietary wire format, nothing to rewrite if you leave.
  ```bash
  curl -X POST https://www.inreels.ai/api/v1/generate_image \
    -H "Authorization: Bearer ir_live_..." \
    -H "Content-Type: application/json" \
    -d '{"prompt":"serum bottle on wet marble","aspect_ratio":"9:16"}'
  ```
- **Bring your own provider keys.** The craft lives in the open
  [inReels skills](https://github.com/inreels-ai/inreels-skills) — the method,
  the prompts, the render scripts. They run on *your* Replicate, fal or
  ElevenLabs key just as happily as on ours. inReels is the default, not a
  requirement.
- **The files are yours.** MP4s and PNGs on a CDN URL you can download and keep.
- **No rate limits** beyond your own credit balance.

You are paying us to not have to hold seven provider accounts, keep up with
model deprecations, and discover for yourself that image-to-video models take
their canvas from the first frame. Not for a cage.

---

## Running it locally

Prefer to keep the key in your own environment, or your client has no remote
MCP support? This repo is a stdio server that forwards to the hosted API, so
the tool list, pricing and validation stay in one place.

```bash
INREELS_API_KEY=... npx -y inreels inreels-mcp
```

Keys: <https://www.inreels.ai/dashboard/settings/api-keys>

---

## The tools

Two layers, deliberately.

**Capabilities** return a finished deliverable and choose the model for you,
exposing creative controls only. **Primitives** render one asset and expose the
full parameter surface, for an agent that wants to drive the models itself.

| Group | Tools |
|---|---|
| Finished deliverables | `create_ad_video` · `create_faceless_video` · `quote_ad_video` · `quote_faceless_video` |
| Single assets, any model | `generate_image` · `generate_video` · `generate_speech` · `generate_sound` · `quote` |
| Assets | `upload_image` · `refresh_asset_url` |
| Discovery | `list_models` · `list_voices` · `list_styles` · `list_actors` |
| Runs and account | `get_run_status` · `list_runs` · `get_account` |

18 tools, 39 models behind them — 8 image, 28 video, plus speech and sound.
FLUX, Nano Banana, GPT Image, Seedance, Kling, MiniMax, Veo, Gemini Omni, Wan,
HeyGen, ElevenLabs.

### How it behaves

- Every generating tool returns a run id immediately. Nothing blocks.
- `get_run_status` with `wait: true` blocks up to 45 seconds and usually
  returns a finished image in one call.
- Video is slower and has two stages — scenes are built, then the MP4 is
  rendered. Roughly 5–10 minutes. A run reports `completed` only once the file
  exists, so an agent should keep polling rather than hand over a watch page as
  the finished thing.
- An unlisted model id is rejected, never quietly substituted.
- Every option is validated against the provider's live schema before it ships,
  so an agent cannot be handed a combination that fails.
- Tools are tagged read, write or sensitive write. Anything that spends credits
  is a sensitive write and is annotated destructive, so your host can ask you
  each time instead of taking blanket permission.

---

## From a terminal

```bash
npx inreels login ir_live_...
npx inreels tools
npx inreels describe create_faceless_video
npx inreels run generate_image --prompt "serum bottle on wet marble" --wait
npx inreels status <run_id>
```

`--json` on any command for scripting. `inreels run` works for any tool,
including ones added after this version was published.

<https://www.inreels.ai/api/v1> returns every tool with its JSON Schema,
unauthenticated. <https://www.inreels.ai/llms.txt> is the same thing written
for an agent to read.

---

## Would rather click than type?

The same engine has a UI at <https://www.inreels.ai> — describe the ad in a
chat box and watch it build. The MCP is for when you want your own agent
driving it.

---

**Links** — [Setup guide](https://www.inreels.ai/connect) ·
[Skills](https://github.com/inreels-ai/inreels-skills) ·
[Pricing](https://www.inreels.ai/pricing) ·
[Support](https://www.inreels.ai/support)

MIT.