Skip to content

Aetelier MCP

Server: atelier-mcp · Transport: stdio · Protocol: Model Context Protocol

The Aetelier MCP server fronts the operational surface of the Aetelier market-data platform as Model Context Protocol tools. Point any MCP client at it and drive the platform in natural language: deploy a market-data collector, get a colocation region recommendation, audit data governance, and fit and serve quantitative models — all from the same chat that already holds your context.

Value

An AI client with this server can carry a task end to end without leaving the conversation. Ask it to "collect Binance BTC and tell me when data is landing" and it deploys the collector, watches health, and reports back. Ask it to "fit a Hawkes model to Kraken ETH trades and forecast the next 50 events" and it lists datasets, runs the fit, and forecasts — adapting when a fit does not converge.

The name is the interface

Each tool does one thing and returns readable JSON. Backend errors come back as JSON ({"ok": false, "error": ...}) rather than protocol errors, so the client observes them and adapts in-loop — for example, loosening the tolerance after a non-converged fit.

Capabilities

Eighteen tools across six domains:

Domain Tools What it does
Platform platform_health Confirm the platform is reachable — call first
Collectors list_collectors · deploy_collector · collector_status · collector_command · collector_health · stop_collector Deploy and manage market-data collectors you own
Deploy prep list_cex_options List the 12-venue option catalog — native symbol formats, market types, orderbook depths, datatype availability — to consult before you deploy
Colocation query_colocation Recommend a deployment region for a venue
Governance governance_catalog · stream_scorecard · governance_datasets Audit stream completeness, quality, and dataset inventory
Fit & serve list_datasets · fit_pricing · fit_hawkes · list_fits · get_fit · forecast_fit Fit a Hawkes model to a dataset, read it, forecast from it

See the tool reference for the per-tool contract and the worked examples for full task flows.

How it fits together

flowchart LR
  U[You] -->|natural language| C[MCP client]
  C -->|MCP over stdio| M[atelier-mcp server]
  M -->|tool calls| P[Aetelier platform]
  P -->|JSON results| M
  M -->|tool results| C
  C -->|answer| U
You ask an MCP client in natural language; the client speaks MCP over stdio to the atelier-mcp server, which calls the Aetelier platform and returns JSON tool results the client reads back to you.

This MCP server is the machine interface to the platform: the same operations a human performs in the webapp, exposed as tools an agent drives.

The server holds no state of its own. Every tool is a thin call into the running platform, and every model, score, and price is computed by the platform — the MCP server is a client surface, not an engine.

Get going

  • Connect a client

    Add the server to Claude Desktop, Claude Code, Cursor, VS Code, or any MCP client.

    Connect →

  • Tool reference

    One fixed record per tool: intent, params, example call and result.

    Read the tools →

  • Worked examples

    End-to-end flows: deploy a collector, watch health, fit and forecast a model.

    See examples →