nable

Cost controls, built into your agents. Ask about your whole cloud and AI bill inside Claude or Cursor, gate what your agents do against your budget, and get the fix as a pull request you approve. One connector to AWS, Azure, GCP and 15 more SaaS and AI providers.

nable is an MCP server that runs on your machine and connects to your real AWS, Azure, GCP, Kubernetes, and SaaS billing. It answers cost questions in your editor, finds waste, drafts the fix for your approval, and verifies the savings landed. Your credentials stay in your OS keychain and your bill never leaves your control, so the no-egress claim is something you can read in the code, not just take on faith. The local agent is open and auditable; a hosted platform is available for teams.

getnable.com · docs, quickstart, and the hosted platform

User: Just downloaded nable: What's our cloud spend this month?

Claude: Here's your spend breakdown:

Total: $8,421.63 / month

Provider    | Spend
------------|----------
AWS         | $7,284.17
Datadog     | $847.20
Snowflake   | $290.26

Top AWS services:
- EC2: $3,184.52
- EKS: $1,821.34
- RDS: $1,243.80
- ElastiCache: $612.40
- S3: $184.22

Most spend is concentrated in us-east-1 ($5,841 of the $7,284 AWS total).

What you can ask

• "What drove our AWS bill up 40% last month?"
• "Which Kubernetes namespace is over-provisioned?"
• "Are there any unusual cost spikes this week?"
• "Which EC2 instances should we downsize?"
• "Compare our cloud spend vs SaaS spend"
• "Create a Jira ticket for any EC2 waste over $200/mo"
• "Which team is spending the most on Datadog?"
• "What will our AWS bill look like next month?"
• "Show me RDS instances with low CPU that we could right-size"
• "What's our effective discount rate from Savings Plans?"

Visual dashboard

finops serve

Starts a local web dashboard your whole team can open in a browser, no Claude Desktop required. Share it with an exec, a FinOps analyst, or anyone who needs to see costs without using an AI interface.

What it shows:

• MTD spend and projected month total
• Cost trend: 3-month historical with run-rate projection
• Efficiency score: composite of waste, commitment coverage, anomaly response, and tag hygiene
• Savings opportunities: ranked by dollar impact, each with a one-click "Mark done" to track actions taken
• Savings pipeline: how much has been identified vs acted on vs verified

The dashboard reads from your local provider connections. Your data stays on your machine.

# Secure with a password (recommended when sharing on a network)
FINOPS_DASHBOARD_PASSWORD=yourpassword finops serve

# Default: auto-generates a random password and prints it at startup
finops serve

Light mode, dark mode, and 30/60/90-day lookback are built in.

Local-first and auditable

Your credentials are encrypted with Fernet and stored in your OS keyring (macOS Keychain, Windows Credential Manager, or libsecret on Linux). They never leave your machine. Cost data is cached in a local SQLite database, and nable has no backend, so we never see your cost data or credentials. One honest caveat: when you ask a question in your AI editor, the figures nable returns go to your editor's own AI to answer it, the same as any prompt, never to a nable server. If you need zero AI exposure, use the local dashboard (finops serve) or CLI, which never touch a model. Teams share findings via Slack alerts, Notion publishing, and CSV exports. No shared database required.

nable is read-only by default. It never writes to your AWS account unless you explicitly enable cleanup mode. Run finops setup aws --iam-template to generate a least-privilege IAM policy with exactly the permissions nable needs.

None of this is take-our-word-for-it. Read the source, check the OpenSSF Scorecard, run finops-doctor to see exactly what nable touches, and set NABLE_NO_TELEMETRY=1 (or FINOPS_AIRGAP=1 to forbid every non-provider request) if you want it locked down.

Manual Claude Desktop config

If finops setup doesn't auto-configure, run:

finops setup claude

Or add manually to claude_desktop_config.json:

With uvx (recommended):

{
  "mcpServers": {
    "nable": { "command": "uvx", "args": ["--python", "3.12", "finops-mcp"] }
  }
}

With absolute path:

{
  "mcpServers": {
    "nable": { "command": "/usr/local/bin/finops-mcp" }
  }
}

Use the path from which finops-mcp.

Config file locations:

• macOS: ~/Library/Application Support/Claude/claude_desktop_config.json
• Windows: %APPDATA%\Claude\claude_desktop_config.json
• Linux: ~/.config/Claude/claude_desktop_config.json

Why uvx? Claude Desktop is a GUI app and doesn't inherit your shell's PATH. uvx sidesteps this by running finops-mcp in its own isolated environment. It's the most reliable option on corporate machines with managed Python installs.

Give your agent cost controls

nable is not just tools your agent reads from. It is a pre-action gate your agent calls before it makes a cost-affecting change: it prices the change, checks it against your budget, and offers a cheaper path. It never applies anything itself. Propose-only, your agent proposes and a human approves.

Add one line to your agent's system prompt (Claude Code, Cursor, or any MCP client):

Before you apply any infrastructure change (a terraform apply, a helm upgrade, creating or resizing a resource) or start an expensive job, first call check_action_policy with the action and the change (a terraform plan, a helm diff, or a monthly_delta_usd). Relay the verdict, the dollar impact, and the cheaper path when one is offered. Never apply a block or an escalate action; surface it to the human. nable is advisory and propose-only.

That turns any agent into a cost-aware one. The gate returns allow / warn / block / escalate against your policy, the monthly and annual dollar impact, a budget verdict labeled with its data age, and a spot alternative when the change is compute. One-way doors (delete, terminate, buy a commitment) and over-budget changes always escalate to a human.

Connectors (17)

Azure roles. The Azure tools span three RBAC roles, granted to the service principal on each subscription. Without them, the affected tools return empty results (run finops doctor to check):

# repeat per subscription
az role assignment create --assignee <client-id> --role 'Cost Management Reader' --scope /subscriptions/<sub-id>
az role assignment create --assignee <client-id> --role Reader --scope /subscriptions/<sub-id>
az role assignment create --assignee <client-id> --role 'Monitoring Reader' --scope /subscriptions/<sub-id>

What nable actually does

nable is not just a connector that pipes billing data into Claude. It runs active analysis on your infrastructure and surfaces findings as tools Claude can reason about and act on.

Every finding is classified by how sure we are. A recommendation is something nable measured: a precise dollar figure, a safe fix, and a check that the savings actually landed on your next bill. An investigation is a signal worth confirming: an honest order-of-magnitude, never a fake-precise number, with the steps to confirm it. nable proposes, you approve, and it verifies. It never changes your infrastructure on its own.

AWS deep audit goes well beyond Cost Explorer. It pulls CloudWatch metrics for every running resource and flags waste that never shows up on your bill: gp2 volumes that should be gp3 (20% cheaper, same performance), unattached EBS volumes, idle NAT Gateways costing $32/mo in base charges, RDS backup retention set way too high, CloudWatch Log Groups with no retention policy growing forever, and Lambda functions allocated 2x the memory they actually use. Think of it as Compute Optimizer plus the layer underneath it.

Anomaly detection uses z-score, CUSUM drift, and day-of-week seasonal normalization. When something spikes, it drills into Cost Explorer by tag and tells you which team, environment, or service drove it. Anomaly findings and Slack/Teams alerts are free; auto-ticketing is a paid feature.

Rightsizing combines AWS Compute Optimizer with nable's own CloudWatch analysis. It gives you specific recommended instance types with estimated savings, not just a list of underutilized resources. Recommendations are free; ticket auto-creation is a paid feature.

Commitment analysis (a paid feature) models Savings Plans and Reserved Instance coverage against your actual usage. It shows your current effective discount rate, coverage gaps, and what you would save by purchasing additional commitments.

Open-core

The local agent is open-source and free: the MCP server, every connector, cost queries, anomaly detection, rightsizing, AI and LLM spend tracking, the local dashboard, and remediation drafts (the PRs and tickets you approve). Run it on your machine, audit it, fork the connectors.

A hosted platform is available for teams who would rather have it run for them: a managed, single-tenant workspace with dashboards anyone can use without a terminal, SSO and roles, scheduled reports, and a managed AI agent. Single-tenant by design, your bill is never pooled with anyone else's.

See getnable.com for the current plans and a free trial.

Docs

Full setup guide: getnable.com/docs

_{mcp-name: io.github.chaandannn/finops-mcp}

Quick start

Requires Python 3.11 or newer. The uvx command below fetches a matching Python for you. If you take the pip path instead, check yours first with python --version (or python3 --version). On older Python, pip reports No matching distribution found for finops-mcp.

Step 1: Install and run the setup wizard

Need uv? It is not preinstalled on macOS or most Linux:

curl -LsSf https://astral.sh/uv/install.sh | sh   # macOS / Linux
# or: brew install uv

Then:

uvx nable

No uv and don't want it? On Python 3.11+, pip install -U finops-mcp && finops welcome works too.

First run downloads dependencies, so give it a moment before the welcome screen appears.

The wizard walks through connecting your providers and auto-configures Claude Desktop at the end. No config file editing, no manual env vars.

Using Cursor? One-click install (opens Cursor and adds nable):

cursor://anysphere.cursor-deeplink/mcp/install?name=nable&config=eyJjb21tYW5kIjogInV2eCIsICJhcmdzIjogWyItLXB5dGhvbiIsICIzLjEyIiwgImZpbm9wcy1tY3AiXX0=

Then run finops setup once to connect a cloud account.

On Anaconda? Use uvx (isolated, won't touch your Anaconda environment):

brew install uv && uvx nable setup

Step 2: Connect AWS (usually one keystroke)

finops setup aws

The wizard checks for AWS credentials you already have (an SSO login, an AWS CLI profile, or default credentials), shows you the account it found, and connects it when you confirm. If you use aws on this machine already, you will not type a single key.

Checking for AWS credentials on this machine...
✓ Found working credentials: profile 'default' -> account 1234
  Connect this account? [Y/n]

Only if no working credentials are found does it walk you through creating a read-only access key. Want the IAM policy to hand your platform team first? Run finops setup aws --iam-template.

Step 3: Restart Claude Desktop and ask

What are my AWS costs this month?

Once you see a real cost breakdown, you're live. Also works with Cursor, Windsurf, and VS Code.

Step 4 (optional): Open the visual dashboard

finops serve

Serves a password-protected web dashboard at http://localhost:8080, local to your machine by default. To let your team or manager view it in a browser (no Claude required), add --host 0.0.0.0 so it binds your network. It stays password-protected; share the URL and password with them.

7-day free trial, all features unlocked. No credit card required.

To add more providers later:

finops setup aws      # add another AWS account
finops setup azure    # add Azure
finops setup slack    # configure alerts
finops setup license  # activate a Team plan key
finops serve          # open the visual dashboard

Troubleshooting

finops-doctor          # checks credentials, DB, network, audit log
finops setup claude    # re-run Claude Desktop configuration only