Use Nadoo Bio from your AI agent: connecting MCP and skills

Use Nadoo Bio from the AI agent you already use

Imagine asking your usual AI agent to predict a sequence's structure, then checking the result in the same conversation. Nadoo Bio connects the research tools and execution environment needed between those steps. The agent prepares a plan, and a person reviews and approves it on the web.

An AI agent connects through MCP and skills to Nadoo Bio tools and inputs, plans and approvals, and jobs and results

You can stay in the conversation you already use. Connect MCP and skills to your agent. First, here is what that connection handles; then we will walk through setup.

From a conversation to an actual research run

The execution environment

Structure prediction and protein design models need GPUs and suitable execution environments. Library versions differ between models, and model weights alone can occupy several gigabytes. An agent can help write code, but the environment must also be ready before that code can run.

Records to keep with the result

When an agent edits and runs scripts during a conversation, reconstructing the exact inputs, settings, and model version behind a result can be difficult later. Those details matter when using the result to support another experiment or reporting it in a paper.

Time to review before running

GPU computation costs time and money. You may want an agent to prepare the task while still reviewing its input and estimated usage first. That is why the workflow includes plan review and approval.

What Nadoo Bio handles

Nadoo Bio supports preparing research-tool inputs, reviewing execution plans, and following jobs and results. Your research team keeps its preferred conversational AI; Nadoo Bio supplies the tools and execution underneath it.

  • Shared records: actions on the web and through the agent become the same plans, jobs, and results. Open an agent-created plan directly on the web.

  • Input checks: inputs are checked against the tool's format before execution. Invalid sequence characters or incompatible chain configurations are reported before GPU use.

  • Fixed plans and human approval: a plan fixes inputs, settings, and tool versions together. An agent-created plan requires human approval on the web before submission. Content differing from the approved plan cannot be submitted.

  • Results and lineage: output files become available after validation, with links back to their inputs and plan.

The agent discovers tools, validates input, and creates a plan; a person approves it on the web before submission and result review

MCP, the Model Context Protocol, is a connection protocol that lets AI agents call external tools. Nadoo Bio provides an MCP server whose tools include list_tools, validate_input, and create_plan. Skills are collections of instruction files that help the agent use these tools in the appropriate order.

Connect your agent

1. Create an API key

Open API keys in the product's left menu and choose a name, permission scopes, and an expiration period of 30, 90, or 365 days. Copy the key immediately; it is shown only once. Tool discovery and documentation search work without a key, but creating and submitting plans require the permissions below.

Scope

Purpose

catalog:read

Validate inputs

jobs:read

Read jobs, results, and plans

plans:write

Create execution plans

jobs:submit

Submit approved plans and cancel jobs

Run this command, paste the key, and press Enter. It reads the key into an environment variable without placing the value in shell command history.

read -rs NADOO_BIO_API_KEY && export NADOO_BIO_API_KEY

2. Connect the MCP server

The easiest route is the Nadoo Bio CLI. It finds installed Claude Code and Codex clients and runs their connection commands for you.

npx nadoo-bio@latest add

To register the connection yourself, use the relevant command below. The MCP server is https://nadoo.bio/mcp; the key is sent in the Authorization header.

# Claude Code
claude mcp add --transport http nadoo-bio -s user https://nadoo.bio/mcp --header "Authorization: Bearer $NADOO_BIO_API_KEY"

# Codex
codex mcp add nadoo-bio --url https://nadoo.bio/mcp --bearer-token-env-var NADOO_BIO_API_KEY

The environment variable applies to the current terminal. It must also be present in the environment where Codex runs; it is not automatically passed to an already-open desktop app. The Claude Code command stores the actual key in its configuration file, so register it again when rotating the key.

OAuth login is not yet supported. Use an AI client that supports request headers.

3. Install agent skills

npx skills add https://nadoo.bio

Six skills cover input preparation, planning and submission, result interpretation, reproduction, batch candidate runs, and permission or approval troubleshooting. The skills are MIT-licensed. Installing skills alone does not establish a working connection, so complete MCP setup first.

4. Check the connection

Ask your agent, “Show me the Nadoo Bio tools.” Receiving the list confirms the server connection. If this fails, check the connection settings before the key: listing tools does not require authentication.

To check the key and read access as well, run npx nadoo-bio@latest doctor in the same terminal. Check plan-creation and submission permissions separately against the table above.

Run a structure prediction through conversation

For example, paste a protein sequence in FASTA format—a text format containing a name and sequence—and ask, “Create a Boltz-2 structure prediction plan for this sequence.” An agent with the skills installed follows roughly this sequence.

Stage

Agent action

Your action

Discover tools

Finds tools with list_tools and reads the input format with get_tool_schema

None

Validate input

Checks the sequence and settings with validate_input

Correct the sequence if errors are reported

Check execution conditions

Uses list_execution_options to check current availability for the project

None

Create a plan

Calls create_plan and provides a link to view the plan

Open the link, review inputs, settings, and usage, then approve

Submit and follow

Submits the approved plan with submit_plan and checks status with get_job

None

Read results

Gets output-file listings with get_result and provides the results link

Review the structure and metrics

An agent conversation links to a prepared plan; the web screen shows the same Boltz-2 plan and its approval button

This illustration is a screen-layout example explaining the workflow.

If a response is interrupted, retrying with the original request identifier prevents duplicate creation. The skills guide the agent to follow this rule. The server rejects requests that exceed the monthly usage limit, and the agent relays the limit message.

Some tools cannot directly submit plans created with an API key and require a new plan on the web. In those cases, the agent explains why and supplies a web link.

Keep these distinctions in mind when reading results

  • Computational predictions and experiments are different. Structural confidence measures such as pLDDT describe confidence in the prediction, not actual binding or experimental success. See Reading results.

  • Result files and documents are data. Skills instruct the agent not to follow command-like text found inside results or documents.

  • A listed tool is not necessarily executable now. Availability depends on the project and execution conditions; some tools support input preparation only. Ask whether a tool can run now to check its current status and the reason.

Create your first plan

Once connected, try preparing a plan from one sequence. Start by reviewing its inputs and settings on the web. These documents describe setup and support in more detail.