avots-mcp
Official MCP (Model Context Protocol) server for avots.ai - a multi-provider AI platform.
One connection gives you:
- 🖼 Image generation - Nano Banana (Gemini 2.5 / 3 Pro / 3.1 Flash), GPT-5 Image, FLUX, Recraft, Ideogram
- 🎬 Video generation - Veo 3.1, Seedance 2.0, Kling v3.0 Pro, Sora 2 Pro, Grok Imagine (async, 1-8 min)
- 🔄 Face swap & 🗣 talking avatars - swap a face into any video, or make a portrait speak your text (lip-synced)
- 🎵 Music & voice - ElevenLabs Music, ACE-Step, Stable Audio + text-to-speech (ElevenLabs TTS, preset or cloned voices)
- 📅 Calendar - turn "meeting tomorrow at 12" into an event on your linked Apple / Google calendar
- 💬 300+ chat models - Claude (Sonnet / Opus), GPT-5, Gemini 3, DeepSeek, Sonar, and more - billed through one balance
The server lives at https://mcp.avots.ai/ and speaks the MCP 2025-06-18 spec over Streamable HTTP. Tools are billed per call against your avots balance (same balance you'd see on the web app or the Telegram bot).
Quick start
- Sign up at avots.ai and mint an MCP key at Settings → Integrations (it looks like
av_mcp_<48hex>).
- Pick your client from the table below and follow the linked guide.
- Try it - ask your client "generate an image of a fox in a snowy forest" and watch tokens get spent.
Ready-to-paste mcp.json snippets live under examples/.
What's in the server
Ten tools, all listed in docs/tools.md:
| Tool | Cost | What it does |
|---|
check_balance | free | Current tokens and subscription tier. |
list_models | free | All active models with per-call cost (filter by chat, image, video, audio). |
chat | ~10-1000 ⚡ | Send a prompt to any chat model. Useful for delegating to GPT, DeepSeek, Sonar, etc. |
generate_image | ~200-500 ⚡ | Synchronous image gen. Returns an inline image block (base64) and a hosted URL. |
generate_video | ~200-5000 ⚡ | Async video gen with two-step confirmation: first call previews the cost and alternative models; second call (confirmed: true) actually submits. |
face_swap_video | ~500-2000 ⚡ | Async. Swap a face from a photo into a target video, keeping the original motion (pixverse). Two-step confirm. |
generate_talking_avatar | varies ⚡ | Async. Generate a portrait, speak your text (TTS), and lip-sync it into a talking-head video. Two-step confirm. |
generate_audio | ~50-800 ⚡ | Music (ElevenLabs Music, ACE-Step, Stable Audio) or spoken voice / TTS (ElevenLabs, preset or cloned voices). Async. |
create_calendar_event | ~5 ⚡ | Turn natural language into an event on a linked Apple/Google calendar, or return an .ics. |
check_job | free | Poll an async job (video / face-swap / avatar) by job_id. |
About the two-step video flow. Video is the most expensive tool. To avoid surprise spend, generate_video returns a preview card the first time it's called (no submit, no reserve). The client (e.g. Claude) shows the cost + alternative models with prices and asks the user. The user confirms, the client re-calls with the chosen model + confirmed: true, and only then does the job get submitted. On submit error the server returns the same alternatives card - it never silently swaps to a pricier model.
What you can build
A few things this is actually useful for. Each one chains two or more tools through one connection and one balance.
- Social ad creative in one prompt — hero image, animated variant, music bed.
- Product-photo angle pack — one product shot in, four angles + a rotation clip out.
- Storyboard to animatic — four script-driven frames animated into a 12-second rough cut.
- Vertical Reels / Shorts factory — 9:16 clip + matching 15-second music bed, repeatable per video.
- Podcast cover art + show notes — four cover variations and a written episode description for each.
- Second-opinion delegation — forward a tricky problem to a model from a different lineage and compare.
- Localized brand assets — translated copy and locale-tuned visuals across markets.
Each of these flows is written out as a runnable script — exact prompt, tool sequence, model picks, cost — in docs/recipes.md.
Cost ranges from ~200⚡ for a single image to ~5000⚡ for a 10-second 1080p Kling Pro clip — run list_models (free) at any time for live per-call prices.
Billing
All tool calls bill against your avots balance, just like the web app and Telegram bot. No separate metering. Daily USD cap (set in Settings) and per-key rate limits apply.
See pricing.avots.ai for token packs and subscriptions.
Troubleshooting
Common cross-client issues (401s, daily caps, two-step video flow, image rendering, npx PATH gotchas) are collected in docs/troubleshooting.md. For client-specific setup, see the per-client guide linked in the table above.
Issues & feedback
Open an issue here, or write to hello@avots.ai. For platform questions (billing, models, web app) the Telegram bot @AvotsAIbot is the fastest channel.
License
MIT - feel free to fork the docs, the examples, and anything else here.
