Skip to main content
Glama

alallo-mcp

A Model Context Protocol server for alallo: fine-tune open language models from an AI agent. Upload training data, get a fixed price before anything starts, start and follow a job, and download the trained model.

Status: preview. API keys are not yet open on alallo.com accounts, and this package is not yet published to npm. Both are coming soon; until then the set-up below won't work.

Set-up

  1. Make an API key in the alallo dashboard: Profile > API keys > Make key. Choose "Prices and data checks" for quotes only, or "Start jobs" to let the agent start and cancel jobs. Set a monthly cap: starts that would go over it are refused.

  2. Add the server to your agent.

Claude Code

claude mcp add alallo -e ALALLO_API_KEY=alk_... -- npx -y alallo-mcp

Cursor (~/.cursor/mcp.json, or .cursor/mcp.json in a project)

{
  "mcpServers": {
    "alallo": {
      "command": "npx",
      "args": ["-y", "alallo-mcp"],
      "env": { "ALALLO_API_KEY": "alk_..." }
    }
  }
}

Codex (~/.codex/config.toml)

[mcp_servers.alallo]
command = "npx"
args = ["-y", "alallo-mcp"]
env = { ALALLO_API_KEY = "alk_..." }

Environment

Variable

ALALLO_API_KEY

your API key (alk_...). Without it only list_models and get_quote without a spec work.

ALALLO_API_URL

https://alallo.com (default; EU accounts, prices in EUR) or https://us.alallo.com (US accounts, prices in USD)

Related MCP server: Tuner MCP

Tools

Tool

What it does

list_models

base models that can be fine-tuned now, and the methods each supports

get_quote

price and time range: fixed for a full job spec, approximate from a data size without one

upload_dataset

uploads a local .jsonl, .json or .csv file and checks it; returns the dataset id

check_dataset

the check result of an uploaded file

start_job

starts a job: without confirm: true it only returns the price; max_price refuses a start above it

get_job, list_jobs

status, progress, price, cost so far, time left

get_job_logs

the last lines of the training log

cancel_job

cancels a job

get_model_download

download links for the trained model and how long it is kept

get_balance

credit available and held, and the page where a person tops up

API keys can't add credit, manage users or make other keys: when the credit runs out the agent gets the top-up page's address to give to a person.

A job spec looks like this (see the API docs):

{
  "name": "support bot",
  "model": "Qwen/Qwen3-0.6B",
  "dataset": { "uploaded": "d0123456789" },
  "training": { "algorithm": "sft-lora", "params": { "max_steps": 100 } }
}

Development

npm install
npm test        # the tools against a mocked API
npm run build   # dist/

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    Related MCP Connectors

    Related MCP Servers

    • A
      license
      C
      quality
      B
      maintenance
      Exposes safe JSON contracts for validating and preparing datasets, running supervised fine-tuning, preference learning, RL, distillation, and evaluation, and managing checkpoints and exports through the Tinker SDK and Cookbook. It lets agents discover models and Hugging Face datasets, plan and execute training recipes, compare experiments, and export selected models.
      48
      Apache 2.0
    • A
      license
      Not graded
      quality
      B
      maintenance
      Enables an AI agent to check live GPU prices, get a cost quote, rent a per-minute GPU instance, upload code and data, run jobs in the background, tail logs, download results and destroy the instance to stop billing. Includes guardrail limits on spend, hours and idle time, safe file transfer, and access to an OpenAI-compatible model API for inference without renting a GPU.
      2,951 npm
      2
      MIT