GetRun MCP Server
Remote Model Context Protocol server for
GetRun — generate images, video, audio and chat with
61 AI models from 16 providers from inside any MCP client.
https://run.getcore.id/mcp
No API key to paste. The server speaks OAuth 2.1 with dynamic client
registration, so a client is authorised once through a consent screen and
billed against the same Rupiah balance as the web app and the REST API.
| |
|---|
| Endpoint | https://run.getcore.id/mcp |
| Transport | Streamable HTTP |
| Auth | OAuth 2.1 + dynamic client registration (RFC 7591) |
| Protocol versions | 2025-06-18, 2025-03-26, 2024-11-05 |
| Scopes | read, generate, manage |
| Tools | 14 |
| Pricing | Per use, in Indonesian Rupiah. No subscription. Failed tasks are not charged. |
This repository holds the connector's public metadata and setup guide. The
service itself is hosted — there is nothing to install and no server to run.
Install
Claude (claude.ai)
Settings → Connectors → Add custom connector. Name it GetRun, paste the
URL above, save, then press Connect and approve the permissions.
Claude Code
claude mcp add --transport http getrun https://run.getcore.id/mcp
Then run /mcp inside a session, pick getrun, and complete the browser
consent flow once.
Cursor
~/.cursor/mcp.json:
{
"mcpServers": {
"getrun": {
"url": "https://run.getcore.id/mcp"
}
}
}
Any other MCP client
Point it at the URL. Clients that only speak stdio can bridge with
mcp-remote:
{
"mcpServers": {
"getrun": {
"command": "npx",
"args": ["-y", "mcp-remote", "https://run.getcore.id/mcp"]
}
}
}
| Tool | What it does |
|---|
getrun_generate_image | Text → image, or image-to-image with a reference |
getrun_generate_video | Text or image → video clip |
getrun_generate_audio | Text → song, sound effect, or speech |
getrun_chat | Chat models, billed per token |
getrun_run_model | Any catalogue model with its raw parameters |
getrun_list_models | Catalogue with prices and JSON Schema per model |
getrun_estimate_price | Cost of a generation before running it |
getrun_get_balance | Available and reserved balance |
getrun_get_task | Status and result of a running generation |
getrun_list_creations | History with prompts, cost and output URLs |
getrun_hide_creation | Tidy the gallery; billing history is kept |
getrun_list_characters | Saved characters — a consistent face across generations |
getrun_save_character | Create or update a character |
getrun_delete_character | Delete a character; past generations remain |
Catalogue
61 models from 16 providers, as of September 2026:
| Category | Models | Examples |
|---|
| Video | 27 | Seedance 2.0, Kling 3.0, PixVerse V6, MiniMax H3 |
| Image | 14 | Nano Banana 2, Seedream 5.0 Pro, GPT Image 2, Qwen3 |
| Text & chat | 14 | Claude Opus 4.8, Claude Sonnet 5, GPT-5.6, Grok 4.6 |
| Audio | 6 | Suno V5, ElevenLabs TTS / Music / Sound Effects |
The live list is always at GET /v1/models —
this table is a snapshot, that endpoint is the source of truth.
Billing
Identical to the REST API: every tool calls /v1 in the same process, so
pricing, balance reservation, schema validation and rate limits are the same,
and results appear in the dashboard's usage history.
Generation tools spend real balance. getrun_estimate_price returns the cost
of a call before you make it, and a task that fails is never charged.
Also available
server.json is the manifest published to the
official MCP Registry under
the DNS-verified namespace id.getrun.
Support
support@getcore.id — include the task_id or request_id when reporting a
failed generation.
Operated by PT Core Digital Asia (GetCore), Indonesia.