Vultr Blade MCP

A GPU-focused MCP server for Vultr cloud infrastructure. 50 tools for VM lifecycle, bare metal, serverless inference, DNS, firewall, snapshots, and billing — with token-efficient defaults and dual safety gates on every write operation.

Implements three service contracts: virtualisation-v1, serverless-v1, dns-authoritative-v1.

Why another Vultr MCP?

| | mcp-vultr | This | |---|---|---| | Scope | 27 service modules (DNS, K8s, VPC, LB...) | GPU compute infrastructure | | Tools | 335+ | 50 (targeted) | | Token cost | Full Vultr API responses | ~60% stripped via formatters | | Write safety | None | Dual gate: env var + per-call confirm | | GPU focus | General-purpose | Plans default to GPU type, VRAM + pricing promoted | | Transport | stdio only | stdio + Streamable HTTP |

mcp-vultr covers the full Vultr platform — DNS, Kubernetes, VPCs, load balancers, block storage, and more. It's comprehensive, but 335+ tool registrations flood the context window before the model reasons about a single VM.

This MCP is designed for agentic platforms that need:

• GPU provisioning, not platform administration — spin up A100s with cloud-init, tear them down when the job finishes.
• Token discipline — formatters strip ~60% of API response for token efficiency. A plan listing shows GPU type, VRAM, count, and hourly price — not 40 fields of metadata.
• Safe writes — every destructive operation requires VULTR_WRITE_ENABLED=true AND confirm: true in the tool call. Creating instances incurs GPU compute charges immediately; accidental invocations are expensive.
• Flexible deployment — stdio for local clients, Streamable HTTP for remote access behind tunnels.

Why "Blade MCP"?

The -blade-mcp suffix identifies this as part of the Blade MCP family — a collection of purpose-built MCP servers that share a common design philosophy:

• Service contracts — each blade implements a declared contract (virtualisation-v1, serverless-v1, dns-authoritative-v1, etc.) so agentic platforms can swap providers without rewriting prompts.
• Token efficiency — formatters strip verbose API responses by default. Full detail on request.
• Dual write gates — environment variable + per-call confirmation on all destructive operations.
• Dual transport — stdio for local use, Streamable HTTP for remote and always-on deployment.

Other blades in the family: cloudflare-blade-mcp (53 tools), caldav-blade-mcp (13 tools), fastmail-blade-mcp (20 tools), and more.

Quick Start

Install

git clone https://github.com/groupthink-dev/vultr-blade-mcp.git
cd vultr-blade-mcp
npm install && npm run build

Configure

# Required — create at my.vultr.com/settings/#settingsapi
export VULTR_API_KEY="your-api-key"

# Required for write operations (create, delete, start, stop, reboot, etc.)
export VULTR_WRITE_ENABLED="true"

Run

# stdio (default — for Claude Code, Claude Desktop)
node dist/index.js

# HTTP transport (for remote access, tunnels)
TRANSPORT=http MCP_API_TOKEN=your-secret node dist/index.js

Claude Code / Claude Desktop

{
  "mcpServers": {
    "vultr": {
      "command": "node",
      "args": ["/path/to/vultr-blade-mcp/dist/index.js"],
      "env": {
        "VULTR_API_KEY": "your-api-key",
        "VULTR_WRITE_ENABLED": "false"
      }
    }
  }
}

Tools (50)

Instances (11 tools)

| Tool | Description | Write | |------|-------------|-------| | vultr_vm_list | List all instances with status, region, GPU, cost | | | vultr_vm_get | Get full details for a single instance | | | vultr_vm_status | Get power state and health (lightweight) | | | vultr_vm_bandwidth | Get bandwidth usage (incoming/outgoing) | | | vultr_vm_create | Create instance with GPU + cloud-init | confirm | | vultr_vm_delete | Permanently destroy an instance | confirm | | vultr_vm_start | Start a stopped instance | confirm | | vultr_vm_stop | Stop (halt) a running instance | confirm | | vultr_vm_reboot | Reboot a running instance | confirm | | vultr_vm_update | Update instance properties (label, plan, tags, firewall) | confirm | | vultr_vm_set_reverse_dns | Set reverse DNS (PTR) for an IPv4 address | confirm |

Bare Metal (9 tools)

| Tool | Description | Write | |------|-------------|-------| | vultr_bm_list | List all bare metal instances | | | vultr_bm_get | Get full details for a bare metal instance | | | vultr_bm_bandwidth | Get bandwidth usage | | | vultr_bm_list_plans | List available bare metal plans with pricing | | | vultr_bm_create | Create bare metal instance | confirm | | vultr_bm_delete | Permanently destroy a bare metal instance | confirm | | vultr_bm_start | Start a stopped bare metal instance | confirm | | vultr_bm_stop | Stop (halt) a running bare metal instance | confirm | | vultr_bm_reboot | Reboot a running bare metal instance | confirm |

Serverless Inference (5 tools)

| Tool | Description | Write | |------|-------------|-------| | vultr_inference_list | List all inference subscriptions | | | vultr_inference_get | Get details for an inference subscription | | | vultr_inference_usage | Get token usage statistics | | | vultr_inference_create | Create an inference subscription | confirm | | vultr_inference_delete | Delete an inference subscription | confirm |

DNS (5 tools)

| Tool | Description | Write | |------|-------------|-------| | vultr_dns_list_domains | List all DNS domains | | | vultr_dns_list_records | List all records for a domain | | | vultr_dns_create_record | Create a DNS record | confirm | | vultr_dns_update_record | Update a DNS record | confirm | | vultr_dns_delete_record | Delete a DNS record | confirm |

Firewall (6 tools)

| Tool | Description | Write | |------|-------------|-------| | vultr_fw_list_groups | List all firewall groups | | | vultr_fw_get_group | Get details for a firewall group | | | vultr_fw_list_rules | List all rules in a firewall group | | | vultr_fw_create_group | Create a firewall group | confirm | | vultr_fw_delete_group | Delete a firewall group | confirm | | vultr_fw_create_rule | Create a firewall rule | confirm |

Snapshots (4 tools)

| Tool | Description | Write | |------|-------------|-------| | vultr_snap_list | List all snapshots | | | vultr_snap_get | Get details for a snapshot | | | vultr_snap_create | Create a snapshot from an instance | confirm | | vultr_snap_delete | Delete a snapshot | confirm |

Startup Scripts (4 tools)

| Tool | Description | Write | |------|-------------|-------| | vultr_script_list | List all startup scripts | | | vultr_script_get | Get details for a startup script | | | vultr_script_create | Create a startup script | confirm | | vultr_script_delete | Delete a startup script | confirm |

Reference Data (4 tools)

| Tool | Description | Write | |------|-------------|-------| | vultr_vm_list_plans | List plans with GPU specs + pricing (default: GPU only) | | | vultr_vm_list_regions | List available datacentres | | | vultr_vm_list_images | List OS images (default: Linux) | | | vultr_vm_ssh_keys | List SSH keys for injection | |

Account (2 tools)

| Tool | Description | Write | |------|-------------|-------| | vultr_account_info | Get account info, balance, pending charges | | | vultr_billing_history | Get billing history with charges and payments | |

GPU Provisioning Flow

1. vultr_vm_list_plans (type=vcg)     → Find GPU plan + pricing
2. vultr_vm_list_regions              → Pick region with availability
3. vultr_vm_list_images               → Get Ubuntu 24.04 os_id
4. vultr_vm_ssh_keys                  → Get SSH key IDs
5. vultr_vm_create                    → Create with cloud-init user_data
   {plan, region, os_id, user_data,      (installs Docker, NVIDIA toolkit,
    script_id, firewall_group_id}         pulls OCI image, starts worker)
6. vultr_vm_status                    → Poll until server_status=ok
7. ... work completes ...
8. vultr_snap_create                  → Snapshot before teardown (optional)
9. vultr_vm_delete                    → Tear down

Token Efficiency

All read tools return compact output by default — only the fields needed for decision-making.

| Resource | Fields returned | Fields stripped | Savings | |----------|----------------|----------------|---------| | Instances | 12 (id, label, status, power, region, plan, IP, cost) | ~28 | ~70% | | Plans | GPU specs promoted (type, VRAM, count) + pricing | metadata bloat | ~60% | | Regions | 5 (id, city, country, continent, options) | ~15 | ~75% | | SSH keys | id, name, date only | key material | ~80% | | Bare metal | 15 (id, label, status, region, plan, CPU, RAM, IP, cost) | ~25 | ~60% | | Inference | id, label, status, masked key | full key, internal | ~70% | | DNS records | 6 (id, type, name, data, TTL, priority) | metadata | ~50% | | Firewall | id, description, counts | metadata | ~60% | | Snapshots | 7 (id, description, status, size, os_id, date) | metadata | ~50% | | Account | 6 (name, email, balance, charges, payment) | ACLs, internal | ~60% |

Responses are capped at 4000 characters with pagination guidance on truncation.

Security Model

| Layer | Mechanism | |-------|-----------| | Environment gate | VULTR_WRITE_ENABLED=true required for any mutation | | Per-call gate | All write tools require confirm: true (Zod z.literal(true)) | | Dual confirmation | Both gates must pass — neither alone is sufficient | | Bearer auth | Optional MCP_API_TOKEN for HTTP transport (timing-safe comparison) | | No credential leakage | API keys never appear in tool responses or error messages | | Inference key masking | Inference subscription API keys are masked in responses | | Cost awareness | vultr_vm_create and vultr_bm_create responses include pricing — GPU/bare metal instances incur charges immediately |

Environment Variables

| Variable | Required | Default | Purpose | |----------|----------|---------|---------| | VULTR_API_KEY | Yes | — | API key | | VULTR_WRITE_ENABLED | For writes | false | Enable create/delete/start/stop/reboot/update | | MCP_API_TOKEN | No | — | Bearer token for HTTP auth | | TRANSPORT | No | stdio | stdio or http | | PORT | No | 8780 | HTTP server port |

Architecture

src/
├── index.ts              — Entry point: stdio / HTTP transport selection
├── server.ts             — MCP server factory, registers all 50 tools
├── constants.ts          — Defaults, env var names
├── tools/
│   ├── instances-read.ts — vultr_vm_list, get, status, bandwidth
│   ├── instances-write.ts— vultr_vm_create, delete, start, stop, reboot, update, set_reverse_dns
│   ├── plans.ts          — vultr_vm_list_plans, list_regions
│   ├── images.ts         — vultr_vm_list_images
│   ├── ssh-keys.ts       — vultr_vm_ssh_keys
│   ├── baremetal-read.ts — vultr_bm_list, get, bandwidth
│   ├── baremetal-write.ts— vultr_bm_create, delete, start, stop, reboot
│   ├── baremetal-plans.ts— vultr_bm_list_plans
│   ├── inference.ts      — vultr_inference_list, get, create, delete, usage
│   ├── dns.ts            — vultr_dns_list_domains, list_records, create, update, delete
│   ├── firewall.ts       — vultr_fw_list_groups, get, create, delete, list_rules, create_rule
│   ├── snapshots.ts      — vultr_snap_list, get, create, delete
│   ├── scripts.ts        — vultr_script_list, get, create, delete
│   └── account.ts        — vultr_account_info, billing_history
├── schemas/              — Zod validation schemas per domain
├── formatters/           — Token-efficient output formatters
├── services/
│   ├── vultr.ts          — API client, key validation
│   └── auth.ts           — Bearer token middleware for HTTP transport
└── utils/
    ├── pagination.ts     — Pagination meta, truncation
    ├── write-gate.ts     — Dual write gate (env + confirm)
    └── errors.ts         — Vultr API error → actionable message mapping

Dependencies: @modelcontextprotocol/sdk, zod. No Vultr SDK — direct REST calls for control over response shaping.

Development

npm install              # install dependencies
npm run dev              # stdio with hot-reload
npm run dev:http         # HTTP with hot-reload on :8780
npm test                 # run tests (vitest, 74 tests)
npm run typecheck        # type-check only
npm run build            # compile to dist/

Stallari Marketplace

This MCP implements three Stallari service contracts (pack-spec 2.0.0):

• virtualisation-v1 — VM + bare metal lifecycle (extended)
• serverless-v1 — inference subscriptions and usage
• dns-authoritative-v1 — domain and record management

It serves as the reference implementation for GPU cloud provider blades.

• Marketplace listing: https://stallari.ai
• Stallari platform repo: https://github.com/Groupthink-dev/stallari
• Plugin manifest: https://github.com/Groupthink-dev/stallari-plugins/blob/main/plugins/tools/vultr-blade-mcp.json

License

MIT