AWS Notebook Runner MCP

Run Jupyter notebooks on temporary AWS compute through an MCP server, with dry-run planning, guardrails, progress reporting, cost estimates, S3 artifacts, and automatic cleanup.

This is not Google Colab automation and it does not bypass provider limits. It is an AI-facing wrapper for AWS notebook execution. The current working execution backend is EC2 + Systems Manager (SSM). A SageMaker Notebook Jobs backend is included for planning and future execution, but it depends on your AWS account quotas.

This project is not affiliated with, endorsed by, or sponsored by Amazon Web Services. AWS and Amazon SageMaker are trademarks of Amazon.com, Inc. or its affiliates.

AI-Readable Manifest

This repository includes manifest.0, a Zero Lang description of the MCP server's purpose, capabilities, and boundaries. It is intended to help AI agents understand that this package is an AWS notebook execution runner with guardrails, not Google Colab automation or a provider-limit bypass tool.

Related MCP server: AWS SSO MCP Server

What It Can Do

• Inspect a local .ipynb under an allowlisted local root.

• Estimate compute cost before launch.

• Build dry-run plans without starting paid compute.

• Start a temporary EC2 instance for a notebook run.

• Execute the notebook through SSM with nbconvert.

• Upload the executed notebook and artifacts to S3.

• Report progress, elapsed time, ETA, SSM status, EC2 state, artifacts, and current compute cost estimate.

• Terminate the EC2 instance automatically after completion.

• Refuse paid compute unless both an environment flag and confirmation token are provided.

What It Does Not Do

• It does not create or broaden IAM permissions.

• It does not manage arbitrary AWS resources.

• It does not open SSH ports.

• It does not provide exact cell-level progress yet.

• It does not include memory/filesystem metrics unless you add SSM snapshots or CloudWatch Agent support.

• It does not make AWS quota requests.

Install

From PyPI, after publication:

pip install "aws-notebook-runner-mcp[aws]"

From a local checkout:

python3 -m venv .venv
.venv/bin/pip install -e ".[aws,test]"

Run the MCP server:

aws-notebook-runner-mcp

Required AWS Resources

You need:

• An S3 bucket/prefix for notebook inputs, outputs, and status files.

• An IAM user or role for the local MCP server.

• An EC2 instance role/profile for temporary notebook instances.

• A default VPC/subnet or explicit subnet id.

• SSM access; no inbound SSH is required.

The tested setup used:

AWS region: eu-north-1
S3 root: s3://YOUR_BUCKET/runs
EC2 instance profile: EC2NotebookRunnerRole
Instance type: t3.micro

IAM: Local MCP User

Attach a managed policy to the IAM principal used by your local AWS profile. Keep it scoped to your account and bucket where possible.

Minimal EC2/SSM/S3 policy shape:

{
  "Version": "2012-10-17",
  "Statement": [
    {
      "Sid": "PassNotebookRunnerRole",
      "Effect": "Allow",
      "Action": "iam:PassRole",
      "Resource": "arn:aws:iam::123456789012:role/EC2NotebookRunnerRole"
    },
    {
      "Sid": "EC2NotebookRunnerControl",
      "Effect": "Allow",
      "Action": [
        "ec2:RunInstances",
        "ec2:TerminateInstances",
        "ec2:CreateTags",
        "ec2:DescribeInstances",
        "ec2:DescribeInstanceStatus",
        "ec2:DescribeImages",
        "ec2:DescribeSubnets",
        "ec2:DescribeVpcs",
        "ec2:DescribeSecurityGroups"
      ],
      "Resource": "*"
    },
    {
      "Sid": "SSMNotebookRunnerControl",
      "Effect": "Allow",
      "Action": [
        "ssm:SendCommand",
        "ssm:GetCommandInvocation",
        "ssm:DescribeInstanceInformation"
      ],
      "Resource": "*"
    },
    {
      "Sid": "NotebookRunnerS3Access",
      "Effect": "Allow",
      "Action": [
        "s3:GetObject",
        "s3:PutObject",
        "s3:DeleteObject",
        "s3:ListBucket"
      ],
      "Resource": [
        "arn:aws:s3:::YOUR_BUCKET",
        "arn:aws:s3:::YOUR_BUCKET/runs/*"
      ]
    },
    {
      "Sid": "OptionalCloudWatchMetrics",
      "Effect": "Allow",
      "Action": "cloudwatch:GetMetricStatistics",
      "Resource": "*"
    }
  ]
}

cloudwatch:GetMetricStatistics is optional. Without it, status still works, but CPU/network/disk I/O metrics are reported as unavailable.

IAM: EC2 Instance Role

Create an EC2 role, for example EC2NotebookRunnerRole, with:

• Trust policy for ec2.amazonaws.com.

• AWS managed policy: AmazonSSMManagedInstanceCore.

• S3 access to the run prefix.

Example inline S3 policy for the EC2 role:

{
  "Version": "2012-10-17",
  "Statement": [
    {
      "Sid": "NotebookRunnerInstanceS3Access",
      "Effect": "Allow",
      "Action": [
        "s3:GetObject",
        "s3:PutObject",
        "s3:DeleteObject"
      ],
      "Resource": "arn:aws:s3:::YOUR_BUCKET/runs/ec2/*"
    }
  ]
}

When you create the role through the AWS console, AWS usually creates an instance profile with the same name as the role.

Environment

Core settings:

export AWS_PROFILE=research
export AWS_REGION=eu-north-1
export AWS_NOTEBOOK_RUNNER_ROOT=/absolute/path/to/local/notebooks
export AWS_NOTEBOOK_S3_ROOT=s3://YOUR_BUCKET/runs
export AWS_NOTEBOOK_BACKEND=ec2_ssm
export AWS_NOTEBOOK_ROLE_ARN=arn:aws:iam::123456789012:role/EC2NotebookRunnerRole
export AWS_NOTEBOOK_ALLOWED_INSTANCE_TYPES=t3.micro,t3.small
export AWS_NOTEBOOK_DEFAULT_INSTANCE_TYPE=t3.micro
export AWS_NOTEBOOK_MAX_RUNTIME_SECONDS=1800
export AWS_NOTEBOOK_MAX_ESTIMATED_COST_USD=1

Paid compute is disabled unless you opt in:

export AWS_NOTEBOOK_RUNNER_ENABLE_EXECUTION=true

The MCP caller must also pass:

confirmation_token = START_PAID_EC2_NOTEBOOK_RUN

SageMaker execution, when quotas are available, uses:

confirmation_token = START_PAID_SAGEMAKER_NOTEBOOK_JOB

MCP Client Configuration

Example stdio config:

{
  "mcpServers": {
    "aws-notebook-runner": {
      "command": "aws-notebook-runner-mcp",
      "env": {
        "AWS_PROFILE": "research",
        "AWS_REGION": "eu-north-1",
        "AWS_NOTEBOOK_RUNNER_ROOT": "/absolute/path/to/notebooks",
        "AWS_NOTEBOOK_S3_ROOT": "s3://YOUR_BUCKET/runs",
        "AWS_NOTEBOOK_BACKEND": "ec2_ssm",
        "AWS_NOTEBOOK_ROLE_ARN": "arn:aws:iam::123456789012:role/EC2NotebookRunnerRole",
        "AWS_NOTEBOOK_ALLOWED_INSTANCE_TYPES": "t3.micro,t3.small",
        "AWS_NOTEBOOK_DEFAULT_INSTANCE_TYPE": "t3.micro",
        "AWS_NOTEBOOK_MAX_RUNTIME_SECONDS": "1800",
        "AWS_NOTEBOOK_MAX_ESTIMATED_COST_USD": "1"
      }
    }
  }
}

Only add AWS_NOTEBOOK_RUNNER_ENABLE_EXECUTION=true when you are ready to allow paid compute, and keep the confirmation token gate.

Tools

• get_runner_status: local policy and dependency status.

• inspect_notebook: validate and summarize a local notebook.

• estimate_notebook_job_cost: estimate SageMaker or EC2 compute cost.

• plan_notebook_job: dry-run SageMaker plan.

• get_sagemaker_notebook_job_spec: return a SageMaker NotebookJobStep spec.

• start_sagemaker_notebook_job: guarded SageMaker execution.

• get_sagemaker_job_status: read SageMaker pipeline execution status.

• check_ec2_setup: read-only EC2/SSM readiness checks.

• plan_ec2_smoke_run: dry-run EC2+SSM plan.

• start_ec2_smoke_run: guarded synchronous EC2+SSM run.

• start_ec2_smoke_run_async: guarded async EC2+SSM run.

• get_ec2_smoke_run_status: EC2/SSM/S3 progress, metrics, artifacts, and cost.

• explain_existing_aws_options: related AWS options and overlap.

Typical EC2 Workflow

1. Inspect the notebook:

inspect_notebook(local_path="notebooks/demo.ipynb")

2. Build a dry-run plan:

plan_ec2_smoke_run(
  local_path="notebooks/demo.ipynb",
  run_name="demo-run",
  instance_type="t3.micro",
  max_runtime_seconds=900,
  instance_profile_name="EC2NotebookRunnerRole"
)

3. Start async execution:

start_ec2_smoke_run_async(
  local_path="notebooks/demo.ipynb",
  run_name="demo-run",
  confirmation_token="START_PAID_EC2_NOTEBOOK_RUN",
  instance_type="t3.micro",
  max_runtime_seconds=900,
  instance_profile_name="EC2NotebookRunnerRole"
)

4. Poll status:

get_ec2_smoke_run_status(run_name="demo-run")

Status includes:

• progress_summary.summary, for example: 70% executing; elapsed wall 4m 44s, compute 4m 42s, ETA 2m 0s

• EC2 instance state.

• SSM command status.

• stdout/stderr tail.

• S3 artifacts.

• Current compute cost estimate.

Progress Model

Progress is phase-based:

created -> staged -> launching -> waiting_ssm -> installing -> executing -> uploading -> completed/failed

This is useful for UX and cost guardrails, but it is not exact cell-level progress. A notebook that sleeps for five minutes will remain in executing until it completes unless the notebook itself writes progress markers.

Cost Model

Cost reporting is an estimate:

• EC2 compute is estimated from instance type, elapsed compute time, and a 60-second minimum.

• EBS, S3 requests/storage, data transfer, and taxes are not included.

• Static prices can be overridden:

export AWS_NOTEBOOK_PRICE_OVERRIDES_JSON='{"t3.micro": 0.0104, "ml.m5.large": 0.115}'

Always verify official costs in AWS Billing or Cost Explorer.

SageMaker Notes

SageMaker Notebook Jobs are the more managed AWS-native way to run notebooks, but new AWS accounts may have a default training-job quota of zero for common instance types. In that case, EC2+SSM is a practical fallback.

The SageMaker backend is included, but EC2+SSM is the path that has been tested end-to-end in this package.

Troubleshooting

iam:PassRole denied:

The local AWS principal needs permission to pass the EC2 instance role:

iam:PassRole on arn:aws:iam::<account-id>:role/EC2NotebookRunnerRole

SSM command never starts:

• Check that the instance role has AmazonSSMManagedInstanceCore.

• Use an Amazon Linux AMI with SSM Agent.

• Ensure the subnet has outbound internet access or VPC endpoints for SSM.

CloudWatch metrics unavailable:

Add cloudwatch:GetMetricStatistics to the local AWS principal.

SageMaker fails with quota zero:

Request quota for the selected instance type, or use the EC2+SSM backend.

Inline IAM policy size exceeded:

Use customer managed policies attached to the user/role instead of adding more inline policies.

Safety

This server is intentionally conservative:

• Dry-run tools do not start compute.

• Execution requires AWS_NOTEBOOK_RUNNER_ENABLE_EXECUTION=true.

• Execution also requires a confirmation token.

• Instance types are allowlisted.

• Max runtime and max estimated cost are policy-controlled.

• EC2 instances are launched with instance-initiated shutdown behavior set to terminate.

Review IAM, S3 prefixes, instance allowlists, and cost caps before enabling execution.