@emfirge/mcp

🛡️ Emfirge

Privacy-first AWS security, inside your AI.

Trace attack paths from the internet to your sensitive data, calculate blast radius, and prove fixes before you apply them — without your resource IDs ever reaching the LLM.

Website · Source · MCP Registry · Privacy · Report an issue

Your AI can read your code, but it can't see your cloud. Emfirge fixes that. It scans your live AWS account, builds a graph of every resource and how they connect, then lets your assistant walk attack paths, simulate breaches, and verify fixes — all from a conversation.

The AI never guesses. Emfirge clones your infrastructure graph, applies the change, and re-runs 58 deterministic rules. Your assistant reads back what the engine proved.

🕸️ Graph-based	Maps every AWS resource and relationship — not isolated resource linting like Checkov/tfsec.
🎯 Attack paths	Weighted-Dijkstra routes from the internet to your data, ranked by exploit difficulty, not hop count.
💥 Blast radius	See exactly what an attacker reaches once they land on a resource.
🔒 Privacy-first	Resource IDs are tokenized on your machine before anything reaches the LLM. The mapping never leaves.
✅ Proven fixes	Clone the graph → apply the change → re-run every rule → diff. A real simulation, not a hunch.
📋 Compliance	CIS AWS Foundations 1.5 + SOC 2, per-control pass/fail, mapped to MITRE ATT&CK.

Install (30 seconds)

npx @emfirge/mcp install

Auto-detects and wires up Claude Desktop, Cursor, Kiro, Cline, Continue, and Codex CLI, then asks you to pick a privacy mode. Restart your client and just ask:

"Scan my AWS account, role arn:aws:iam::123456789012:role/EmfirgeReadOnly, region us-east-1"

No role yet? Say "help me set up Emfirge" — your assistant hands you a one-click CloudFormation deploy link for a read-only IAM role.

Free. 5 scans/day per AWS account. No signup. No API keys.

Try it with zero setup

Use the demo ARN — fake infrastructure, the real engine:

arn:aws:iam::194722410583:role/EmfirgeReadOnly    region: us-east-1

"Scan with arn:aws:iam::194722410583:role/EmfirgeReadOnly in us-east-1"

Want a visual graph instead? emfirge.cloud — same engine, browser UI, free during beta.

What your AI gets

Tool	What it does
emfirge_setup_help	Returns a clickable CloudFormation deploy link (for first-time setup).
emfirge_scan	Scan an AWS account — returns risk score, finding counts, and an analysis_id.
emfirge_get_findings	Full findings list for a scan, filterable by severity.
emfirge_attack_paths	Attack paths from the internet to internal resources, plus chokepoints.
emfirge_verify_fix	Simulate a fix and see the real score delta — no changes to your AWS.
emfirge_check_compliance	CIS AWS Foundations / SOC 2 per-control status.
emfirge_simulate_breach	Full kill-chain walkthrough — attack stages, blast radius, follow-up moves.

All seven tools are deterministic on the backend — no LLM calls inside the MCP path. Your host LLM (Claude / Cursor / etc.) is the only AI in the loop, and in strict mode it only ever sees tokenized data.

Privacy by default

In strict mode (the default), every AWS identifier is tokenized locally before it reaches your LLM:

What the LLM sees:    "SG_001 has SSH open → reaches S3_001"
What's on your disk:  SG_001 = sg-0a1b2c3d
                      S3_001 = acme-customer-pii

The mapping lives at ~/.emfirge/tokens.json and is never sent to Emfirge, Anthropic, or anyone. When you say "fix SG_001", the MCP resolves the real ID locally, calls the backend, and re-tokenizes the response.

Mode	What's tokenized	Best for
strict (default)	Every AWS ID — ARNs, EC2/SG/IAM/S3, IPs, account IDs, bucket names	Banks, healthcare, regulated industries
balanced	ARNs, EC2/SG/EIP/IAM IDs, IPs, account IDs. Subnets/VPCs/volumes raw.	Most users
off	Nothing — raw IDs go to the LLM	Personal accounts, demo, debugging

npx @emfirge/mcp privacy strict|balanced|off   # change mode across every wired client
npx @emfirge/mcp privacy                        # show current mode

Honest note: tokenization sits between the MCP and the LLM. The Emfirge backend does receive real IDs — it has to, to call AWS. It stores them for 90 days, then auto-deletes. Full details in PRIVACY.md. Wipe everything anytime with npx @emfirge/mcp purge --role-arn <ARN>.

How it works

┌──────────────┐   role ARN    ┌──────────────┐  read-only   ┌─────┐
│ Your machine │──────────────▶│ emfirge.cloud│─────────────▶│ AWS │
│  (MCP host)  │               │   (scanner)  │  STS, 1 hr   └─────┘
└──────┬───────┘               └──────┬───────┘
       │ tokenized IDs                │ findings + graph
       ▼                              ▼
┌──────────────┐               ┌──────────────┐
│  Your LLM    │               │ Postgres + S3│
│ (Claude/etc) │               │  (90-day TTL)│
└──────────────┘               └──────────────┘

• You ask your AI to scan. The MCP calls the backend with your read-only role ARN.
• The backend assumes the role (1-hour STS token, ExternalId-scoped), scans 16 AWS services, builds the graph, runs the rules, and returns findings.
• The MCP tokenizes every resource ID locally, then hands the safe version to your LLM.
• Your assistant reasons over it — attack paths, fixes, compliance — and you never leave the chat.

Under the hood

• Weighted Dijkstra attack paths — edges weighted by exploit difficulty (0 = metadata, 1 = trivial network reach, 3 = needs a shell). A 5-hop trivial path ranks more dangerous than a 2-hop credential theft.
• Brandes' betweenness centrality — finds chokepoint nodes where hardening one resource kills the most attack paths at once.
• Deterministic fix simulator — graph mutation + full rule re-run, no LLM in the verification path. Proof, not a guess.
• 58 graph-aware rules with context-aware severity — SSH-open behind an ALB drops Critical → Low; public S3 with CloudFront drops Critical → Low.
• Toxic-combo detection — dangerous pattern pairs like public RDS + no CloudTrail.

Coverage: EC2, Lambda, ECS, S3, EBS, RDS, IAM, Secrets Manager, KMS, VPC, Security Groups, WAF, CloudFront, SNS, CloudTrail, GuardDuty, CloudWatch, AWS Config, Budgets.

Manual configuration

If auto-install doesn't work, add this to your client's MCP config:

{
  "mcpServers": {
    "emfirge": {
      "command": "npx",
      "args": ["-y", "@emfirge/mcp"],
      "env": { "EMFIRGE_PRIVACY": "strict" }
    }
  }
}

Config file locations:

• Claude Desktop — ~/Library/Application Support/Claude/claude_desktop_config.json (Mac), %APPDATA%\Claude\claude_desktop_config.json (Windows)
• Cursor — ~/.cursor/mcp.json
• Kiro — ~/.kiro/settings/mcp.json
• Cline — ~/Library/Application Support/Cline/cline_mcp_settings.json
• Continue — ~/.continue/config.json

MCP requires a desktop AI client (stdio transport). Web Claude / ChatGPT / Gemini don't support MCP yet — use emfirge.cloud for those.

CLI

npx @emfirge/mcp install                         # auto-wire to all detected clients
npx @emfirge/mcp install --privacy=balanced      # non-interactive: skip the prompt
npx @emfirge/mcp uninstall                        # remove from all clients
npx @emfirge/mcp status                           # show what's wired up + privacy mode
npx @emfirge/mcp privacy <strict|balanced|off>    # change privacy mode everywhere
npx @emfirge/mcp tokens                           # list local token mappings
npx @emfirge/mcp purge --role-arn <ARN>           # delete all your scan data

Environment variables

Variable	Default	Purpose
EMFIRGE_BASE_URL	https://emfirge.cloud/api	Backend URL — override to point at a self-hosted backend
EMFIRGE_PRIVACY	strict	strict, balanced, or off
EMFIRGE_TRUSTED_ACCOUNT_ID	282027772803	AWS account ID to trust in the IAM role (for setup_help)
EMFIRGE_EXTERNAL_ID	aws-risk-agent	ExternalId for STS assume-role

Security model

• Read-only IAM role — zero write permissions.
• ExternalId — prevents confused-deputy attacks.
• Scoped trust — only Emfirge's AWS account can assume the role.
• STS temporary credentials — expire in 1 hour, never stored.
• Instant revoke — delete the CloudFormation stack and all access is gone.

License

BUSL 1.1 — free for non-production and small production use (up to $1M ARR or 100 employees). Auto-converts to Apache 2.0 in 2030.

See your cloud the way an attacker does — from inside your AI.

Website · Source · MCP Registry