StackServe MCP
Workload-aware cloud stack advisor. An MCP server that recommends complete, compatible cloud stacks with line-item decimal cost estimates (rust_decimal), caveats, a cheaper baseline and a performance-oriented alternative, plus project-specific setup bundles.
Tools
| Tool | Purpose |
|---|---|
profile_project | Extract runtime, database, storage, and unresolved questions from a project brief or repo |
search_provider_docs | Retrieve cited official provider doc chunks with source_url and verified_at |
recommend_stacks | Filter incompatible services and rank feasible stacks by cost and operational trade-offs |
compare_stacks | Compare itemized decimal costs across current, 3x traffic, and 10x storage/egress scenarios |
generate_setup_bundle | Emit .env.example, DEPLOYMENT.md, architecture-decision.md, provisioning-prompt.md, and resource-manifest.json |
Quick Connect (Remote MCP)
Connect directly to the hosted Streamable HTTP endpoint (https://mcp.stackserve.net/mcp) with zero local build:
Claude Code
claude mcp add --scope user --transport http stackserve https://mcp.stackserve.net/mcp
Or install as a plugin:
/plugin marketplace add kztJames01/StackServe
/plugin install stackserve@stackserve-marketplace
OpenAI Codex CLI plugin
codex plugin marketplace add kztJames01/StackServe
codex plugin add stackserve@stackserve-marketplace
Start a new CLI session after installation.
Codex VS Code extension (MCP)
The installed Codex extension can expose the OpenAI plugin directory, as shown in the owner's screenshots. Directory publication is separate from installing this repository through the CLI marketplace. Until StackServe is approved and available in that directory, connect its five tools directly through MCP:
codex mcp add stackserve --url https://mcp.stackserve.net/mcp
codex mcp login stackserve
Alternatively, merge this entry into ~/.codex/config.toml, then complete OAuth sign-in:
[mcp_servers.stackserve]
url = "https://mcp.stackserve.net/mcp"
Cursor plugin and MCP
The root plugin.json, mcp.json, and skills/ form a portable Agent Plugin supported by Cursor. Public listing in Cursor's marketplace requires a separate submission and review. See client distribution for the submission path and verification status.
For direct MCP access before marketplace approval:
Add to .cursor/mcp.json (or ~/.cursor/mcp.json):
{
"mcpServers": {
"stackserve": {
"url": "https://mcp.stackserve.net/mcp"
}
}
}
OpenCode & Kilo
For Kilo, merge the following into project .kilo/kilo.json or global ~/.config/kilo/kilo.json. For OpenCode, use opencode.json:
{
"mcp": {
"stackserve": {
"type": "remote",
"url": "https://mcp.stackserve.net/mcp"
}
}
}
Authentication
The hosted endpoint (https://mcp.stackserve.net/mcp) is protected by Cloudflare Access with managed OAuth. Add the URL to your client as usual: on first use the client opens a browser sign-in and then stores the session. There is no token to copy or rotate.
- Clients with MCP OAuth support (Claude Code, Cursor, Inspector, ChatGPT connectors) need no extra configuration.
- Clients that cannot complete an OAuth flow can use an Access service token (
CF-Access-Client-IdandCF-Access-Client-Secretheaders) issued by the maintainer, or run the server locally. /healthzstays unauthenticated so platform health checks keep working.
Self-hosted HTTP fails closed: serve-http will not start without a bearer token.
STACKSERVE_HTTP_BEARER_TOKEN="$(openssl rand -hex 32)" stackserve-mcp serve-http --bind 127.0.0.1:8787
Clients then send Authorization: Bearer <token> to /mcp. Set STACKSERVE_HTTP_ALLOWED_HOSTS (comma-separated) when serving behind a domain the server does not already allow.
Offline Local CLI & Docker
Run completely offline without network access:
# Docker (offline stdio)
docker run --rm -i --network none ghcr.io/kztjames01/stackserve-mcp:0.1.0
# Or install the CLI onto $PATH
cargo install --git https://github.com/kztJames01/StackServe stackserve-mcp
stackserve-mcp recommend workload.json
Privacy & limits
- StackServe does not provision resources, collect telemetry, or require cloud credentials. Never supply real secrets as project evidence.
- Remote MCP calls send their supplied arguments to the hosted service. Local offline execution and hosted execution have different privacy boundaries; infrastructure logging and retention must be covered by the hosted service's published privacy policy.
- Estimates use a pinned catalog snapshot—always verify each line item's
source_urlandchecked_atagainst official provider pricing and terms before purchase.
Chat plugin preparation
plugin.json and mcp.json provide portable Agent Plugins source for the existing five tools. skills/stackserve/SKILL.md describes chat usage; a custom embedded UI is not required. The initial intended listing is free, by Kaung Zaw Thant, in all available countries.
Repository installation and public-directory approval are separate. See client distribution for installation, marketplace submission, and verification status. The OpenAI public-directory listing remains incomplete. Only package the manifests, referenced assets, and skills/stackserve/; do not archive the repository, local configuration, model artifacts, or dependencies. Existing client-specific manifests remain for their respective clients.
License
MIT OR Apache-2.0. Copyright (c) 2026 Kaung Zaw Thant.