What nable is, in one minute

nable answers questions about your cloud and AI bill right inside Claude or Cursor. Ask why a bill went up, it finds the cause, drafts the fix, and proves the savings. No SQL, no dashboards.

An engineer installs nable and connects the read-only access once. The IAM, CloudFormation, and SSO pages below are for them. After that, anyone on the team just asks questions, no setup, no terminal.

Quickstart

Install nable, point it at your AI editor, then ask about your cloud spend. You only need one provider to start. Everything stays on your machine.

uvx nable

uv tool install finops-mcp && uv tool update-shell

// Claude Desktop: ~/Library/Application Support/Claude/claude_desktop_config.json
// Cursor: ~/.cursor/mcp.json
{
  "mcpServers": {
    "finops": { "command": "finops-mcp" }
  }
}

Detailed installation

The long version of the quickstart, with verification and the manual MCP path spelled out. Takes about five minutes including IAM setup.

uvx nable

# Scans this machine for credentials (env vars, gcloud, gh, ~/.modal.toml…)
# and connects everything it finds in one keystroke:
finops connect
# Or pick providers individually:
finops setup aws
finops setup gcp
finops setup azure

{
  "mcpServers": {
    "finops": { "command": "finops-mcp" }
  }
}

finops doctor
✓ aws      configured
✓ datadog  configured
– azure    not configured

Activate your Team license

After purchase, your license key is emailed to you. Keys start with FINOPS-2-. Run one command and it handles everything.

finops setup license FINOPS-2-your-key-here

AI client setup

Add the MCP server to your AI client's config. You only need to do this once. If you ran finops setup, it was configured automatically with the correct path.

{
  "mcpServers": {
    "finops": { "command": "finops-mcp" }
  }
}

{ "mcpServers": { "finops": { "command": "finops-mcp" } } }

{ "mcpServers": { "finops": { "command": "finops-mcp" } } }

finops-mcp  # listens on stdio

AWS

Connects to Cost Explorer for spend data and CloudWatch for rightsizing metrics. Supports IAM keys, IAM Identity Center SSO, and cross-account role assumption.

Permission tiers

nable works with whatever you're comfortable giving it. Start with Cost Explorer read-only for cost queries, anomaly detection, and budgets. Add the audit permissions when you're ready for the deep waste scanner.

ce:GetCostAndUsage
ce:GetSavingsPlans*
ce:GetReservation*
sts:GetCallerIdentity

# everything on the left, plus:
cloudwatch:GetMetricData
ec2:Describe{Instances,Volumes,
  Snapshots,Addresses,NatGateways}
rds:Describe{DBInstances,DBSnapshots}
lambda:ListFunctions
logs:DescribeLogGroups
s3:ListAllMyBuckets
compute-optimizer:Get*

Setup

finops setup aws

Manual env vars

AWS_ACCESS_KEY_ID=AKIA...
AWS_SECRET_ACCESS_KEY=your-secret
AWS_DEFAULT_REGION=us-east-1  # CE requires us-east-1

Azure

Connects to the Azure Cost Management API. Supports a Service Principal (recommended for automation) and device-code flow (interactive login).

Setup

az ad sp create-for-rbac --name finops-mcp \
  --role "Cost Management Reader" \
  --scopes /subscriptions/YOUR_SUB_ID

finops setup azure

Manual env vars

AZURE_TENANT_ID=your-tenant-id
AZURE_CLIENT_ID=your-client-id
AZURE_CLIENT_SECRET=your-client-secret
AZURE_SUBSCRIPTION_ID=your-subscription-id

GCP

Primary path: BigQuery billing export (recommended, richer data). Fallback: Cloud Billing API. Both use a service-account JSON key stored encrypted in your vault.

Setup

finops setup gcp

Manual env vars

GOOGLE_APPLICATION_CREDENTIALS=/path/to/key.json
GCP_PROJECT_ID=my-project-123
GCP_BIGQUERY_DATASET=billing_export  # if using BQ export
GCP_PROJECT_IDS=my-prod,my-staging  # enables the GCP waste audit

Cost guardrail for agents

One command wires nable into an AI coding agent so it checks every infrastructure-mutating command against your policy before it runs. No prompt engineering. It is advisory and propose-only: nable pauses or blocks an action and shows why, but it never executes anything itself. Read the agent cost controls overview for the why.

Install

In your project (or add --global for every repo), run:

finops guard install

This writes a Claude Code PreToolUse hook to .claude/settings.json. Restart Claude Code so it picks up the hook. Uses uvx nable? The hook is finops guard hook either way. Check status with finops guard status, remove with finops guard uninstall.

What it does

Before your agent runs an infrastructure command, nable classifies it and applies your policy:

Try any command against your policy without an agent:

finops guard check --command "terraform destroy"

Policy knobs

Safety

The guard never executes a command; it only returns ask, deny, or nothing. It fails open, any internal error exits cleanly so a guard bug can never break your agent, and it adds about 80ms per command. Propose-only stays fully intact: nable does not auto-execute infrastructure. Outside Claude Code (Cursor, Windsurf, your own loop), point the agent at the same gate with a one-line system-prompt instruction, see the adoption paths.

Security & permissions

nable is not SaaS. It runs on your machine, holds credentials in your OS keyring, queries your cloud providers directly, and has no backend that receives your data. It reads your cloud bill. It cannot change your cloud. Full detail on the security page and the access scope page.

What nable can and cannot do

There is exactly one optional write permission in the whole footprint: logs:PutRetentionPolicy, used only if you explicitly ask nable to fix a CloudWatch Log Group that has no retention policy. You can omit it and lose nothing but that one auto-fix. Destructive cleanup is off entirely unless you set FINOPS_CLEANUP_ENABLED=true yourself.

Where things live

Verify before you trust

You do not have to take this on faith. nable ships the tools to generate a scoped credential and prove it is read-only:

Because nable runs entirely on your machine with no backend, you can watch the wire yourself: run it behind a packet capture (tcpdump, Little Snitch, your egress firewall) and confirm the only outbound destinations are your own cloud provider APIs. Two exceptions, both yours to disable: anonymous opt-out usage telemetry (no cost figures, account IDs, or credentials; off with NABLE_NO_TELEMETRY=1); and, only if you opt into peer benchmarking which is off by default, anonymized differentially-private cost ratios. FINOPS_AIRGAP=1 forbids all non-provider traffic including both.

AI exposure, stated plainly

When you ask a question in Claude, Cursor or another MCP editor, the cost figures nable returns are sent to that editor's AI model to compose the answer, the same as anything else in the chat. nable itself sends nothing to any model. For zero model exposure, use the local dashboard or CLI, which never touch one. Enterprises routing AI through private endpoints (Bedrock, Azure OpenAI) keep that boundary too.

Team mode

Least-privilege IAM

The exact, least-privilege read-only permissions nable needs on each provider. Hand the policy below to your platform team, or run finops setup aws --iam-template to generate it. Every action is a Get, List, Describe, or read.

AWS least-privilege read-only policy

{
  "Version": "2012-10-17",
  "Statement": [{
    "Sid": "NableReadOnlyFinOps",
    "Effect": "Allow",
    "Action": [
      "ce:GetCostAndUsage", "ce:GetCostForecast",
      "ce:GetReservationUtilization", "ce:GetReservationCoverage",
      "ce:GetSavingsPlansPurchaseRecommendation",
      "ce:GetSavingsPlansUtilization", "ce:GetSavingsPlansCoverage",
      "ce:GetRightsizingRecommendation", "ce:ListCostAllocationTags",
      "compute-optimizer:GetEC2InstanceRecommendations",
      "compute-optimizer:GetLambdaFunctionRecommendations",
      "compute-optimizer:GetRDSDatabaseRecommendations",
      "compute-optimizer:GetECSServiceRecommendations",
      "compute-optimizer:GetEnrollmentStatus",
      "cloudwatch:GetMetricData", "cloudwatch:GetMetricStatistics",
      "cloudwatch:ListMetrics",
      "ec2:DescribeInstances", "ec2:DescribeRegions", "ec2:DescribeVolumes",
      "ec2:DescribeSnapshots", "ec2:DescribeAddresses",
      "ec2:DescribeNatGateways", "ec2:DescribeImages",
      "elasticloadbalancing:DescribeLoadBalancers",
      "rds:DescribeDBInstances", "rds:DescribeDBSnapshots",
      "lambda:ListFunctions", "lambda:GetFunctionConfiguration",
      "ecr:DescribeRepositories", "ecr:DescribeImages",
      "ecs:ListClusters", "ecs:ListServices",
      "ecs:DescribeServices", "ecs:DescribeTaskDefinition",
      "logs:DescribeLogGroups", "logs:DescribeLogStreams",
      "logs:PutRetentionPolicy",  // the one optional write, omit to be pure read-only
      "cloudtrail:DescribeTrails", "cloudtrail:GetTrailStatus",
      "cloudtrail:GetEventSelectors",
      "s3:ListAllMyBuckets", "s3:GetBucketLocation",
      "s3:GetBucketIntelligentTieringConfiguration",
      "s3:ListBucketMultipartUploads",
      "organizations:ListAccounts", "organizations:DescribeAccount",
      "organizations:DescribeOrganization",
      "sts:GetCallerIdentity", "sts:AssumeRole"
    ],
    "Resource": "*"
  }]
}

Every action is a Get, List, Describe, or read. There is no Create, Delete, Modify, Run, Terminate, or Put anywhere except the single annotated log-retention line.

One optional, separate backend: if you turn on the detailed CUR-via-Athena cost path (off by default, not part of this policy), it also needs read access to Athena and Glue plus a scoped s3:PutObject to its own Athena query-results bucket. That is the only write outside log retention, it is opt-in, and it is scoped to a bucket you designate.

Google Cloud roles to grant the service account

All are read and query roles. nable reads your Cloud Billing export in BigQuery and, for the optional waste audit, lists Compute resources and reads Monitoring metrics. It writes to nothing. Set GCP_PROJECT_IDS to turn the resource audit on.

Azure RBAC roles per subscription

All three are Reader-tier roles. nable performs no write operation on Azure.

Enterprise SSO (OIDC)Enterprise

Lets your whole team sign in with their company credentials via Okta, Azure AD, Google Workspace, Auth0, or any OIDC-compatible IdP. IdP groups automatically map to nable RBAC roles, with no manual API key distribution.

Setup

finops setup sso

Role mapping example

OIDC_ISSUER=https://yourcompany.okta.com
OIDC_CLIENT_ID=0oa1abc...
OIDC_CLIENT_SECRET=your-client-secret
OIDC_GROUPS_CLAIM=groups
OIDC_ROLE_MAP={"finops-admins":"admin","engineering":"analyst","finance":"viewer"}
OIDC_DEFAULT_ROLE=viewer  # fallback for users not in any mapped group

RBAC roles

All environment variables

Every variable nable reads. The setup wizard writes these for you; this table is for manual config, CI, and containerized runs.

Troubleshooting

The usual suspects, and how to clear them fast.

# Find where pip puts executables
python3 -m site --user-base
# Add /bin to that path, e.g.:
export PATH="$HOME/Library/Python/3.11/bin:$PATH"

Next steps