Use Denario with Claude

Denario runs as a hosted MCP server. You install a small client on your machine, point it at the server with an API key, and the research tools appear inside Claude. The private pipeline runs on our infrastructure; your project files and Git repo are created locally, under your own identity.

Prerequisites

You'll need:

Python 3.10+
Claude — Claude Code or Claude Desktop

Optional — version control. If git and the GitHub CLI (gh) are installed and signed in (gh auth login), Denario also commits your project as it goes and creates a repo under your own account. Without them the project is still created and every file is written to your machine as usual.

Get an API key

Sign in and generate a key on the dashboard. You'll get a key that looks like denario_live_…. Copy it immediately — it's only shown once.

Create an API key →

Install the client

The client is published on PyPI:

bash

pip install denario-mcp

It's intentionally lightweight (it only depends on mcp and httpx) and contains none of the research pipeline — that stays on the server.

Already installed? Upgrade to get the latest tools:

bash

pip install -U denario-mcp

Add Denario to Claude

Claude Code — one command (recommended):

Run this in your terminal with the key you just created. No file editing needed:

bash

claude mcp add denario \
  -e DENARIO_API_KEY=denario_live_your_key_here \
  -e DENARIO_CLOUD_URL=https://denario-mcp-engine-tnf7nmfboa-uc.a.run.app \
  -- denario-mcp

-e sets each environment variable, and -- separates the launch command. By default the server is added for the current project only; add -s user to make it available in all your projects. Check it registered with claude mcp list, or manage it in-session with the /mcp command.

Claude Desktop (or editing config by hand):

Add this block to your MCP servers configuration — for Claude Code that file is ~/.claude.json (or a project .mcp.json):

json

{
  "mcpServers": {
    "denario": {
      "command": "denario-mcp",
      "env": {
        "DENARIO_API_KEY": "denario_live_your_key_here",
        "DENARIO_CLOUD_URL": "https://denario-mcp-engine-tnf7nmfboa-uc.a.run.app"
      }
    }
  }
}

Restart Claude so it picks up the new server.

Run your first project

Just ask Claude in natural language. Two tools are available today — denario_setup (runs locally) and denario_idea (private compute on the server, written back to your project). See the MCP Endpoints page for details.

Initialize a project:

Use denario_setup to create a project at ~/research/oscillators
with this data description: "Damped harmonic oscillator time series
at 1 kHz, 50 trials; infer damping from noisy traces and test
whether temperature biases the result."

This creates the project folder and its data description on your machine.

Generate a research idea:

Now run denario_idea on ~/research/oscillators.

The pipeline runs on the server (~2–3 minutes); the resulting idea.md is written into your project and committed automatically.

Optional settings

Add these to the env block of the MCP config. Neither is required to run Denario:

Variable	Purpose
DENARIO_GITHUB_PUSH	Off by default: your project is a local git repo and nothing is pushed to GitHub. Set to `1` to opt in — Denario then creates a repo under your own identity and pushes after each step.
GITHUB_ORG	Only if you enable the GitHub integration above: the account/org the project repo is created under.
SCIENTIST_NAME	Used in the derived repo name and README.

Troubleshooting

401 Unauthorized: Check DENARIO_API_KEY is a valid, active key from the dashboard.
503 the key checker is unavailable: Nothing to do with your key — the service that checks it could not be reached. Try again shortly.
402 spending limit reached: Your key is valid. This month's budget for you or your organisation has run out — an administrator can raise it on the organisation page.
git or gh errors: Harmless — these only come from the optional version-control step. Your project and every generated file are still written to your machine.
denario_idea takes a while: It's running the real pipeline (~2–3 minutes) — this is expected.