mcp-onprem-starter

A starter template for building single-client MCP servers that expose an existing REST API through MCP. The primary deployment profile runs as a stdio subprocess in a customer-controlled environment, including private cloud, data-center, and air-gapped infrastructure.

Architecture

MCP client / agent
        │
        │ stdio
        ▼
MCP server
  ├── tool handlers
  ├── write authorization
  ├── typed errors
  └── HTTP client
        │
        ▼
Configured upstream API

Development note: When UPSTREAM_BASE_URL is unset, the server uses the bundled mock upstream so the request path can run locally without an external API.

Related MCP server: MCP Server Templates

Quick start

npm install
npm run dev

For a reproducible development environment, use the optional Nix flake:

nix develop
npm ci
just verify

The development shell provides Node.js, just, and the Docker CLI. Nix is a developer-environment option; the server's runtime and package installation remain based on Node.js and npm.

This starts the server over stdio with a bundled mock upstream and the development defaults. Point an MCP client (Claude Code, an agent runtime, or the SDK's own test client) at it and call list_records or create_record.

Template note: list_records, create_record, and the bundled mock are illustrative examples. Replace them with tools and upstream mappings derived from the operator's business request.

To verify the whole deployable chain — build, container boot, a real protocol round trip, and a rollback drill — run:

just verify

Docker is the default container runtime. Podman is also supported:

CONTAINER=podman just verify

What's included

Implementing an operator request

Agents implementing a business request should first identify the required workflows, upstream API contract, tool behavior, security requirements, and acceptance criteria. See AGENTS.md for the complete implementation workflow and repository rules.

Adapting this template

Use this sequence to build the business-specific server:

1. Understand the structure. Read src/index.ts, src/config.ts, src/http/client.ts, src/tools/shared.ts, and src/errors.ts.

2. Define the upstream contract. Identify the API base URL, authentication method, endpoints, request and response schemas, and error behavior.

3. Design the MCP tools. Map business operations to tools with clear names, descriptions, input schemas, annotations, and result shapes.

4. Implement the tool handlers. Add one module per tool under src/tools/, use the shared HTTP client, apply checkWriteGate() to writes, and register each tool in src/tools/index.ts.

5. Add configuration. Extend ConfigSchema for required URLs, credentials, timeouts, or capability settings. Keep environment access inside src/config.ts.

6. Add tests. Cover tool inputs, upstream responses, error mapping, write authorization, and the MCP initialize/tool-listing round trip.

7. Verify the deployment. Run npm run ci, then just verify for the container, non-root boot, stdio, and rollback checks.

Template cleanup

After the business tools are implemented:

• Replace src/tools/_example-read.ts and src/tools/_example-write.ts.

• Update the bundled mock and its tests to represent the business domain, or retain them as a local contract-test harness.

• Remove placeholder endpoint definitions and example terminology.

• Update .env.example, README.md, and DEPLOYMENT.md.

• Keep the shared configuration, HTTP client, error taxonomy, write gate, logging, and verification structure when they still match the deployment.

Design principles

• stdio by default. The server exposes the MCP connection through stdio. The agent runtime spawns this as a child process inside its own trust boundary.

• Validate at startup, isolate optional capabilities. A malformed required setting stops startup with the variable name in the message. A missing optional credential disables just the tool that needs it — the server still starts, and the disabled tool stays listed rather than silently vanishing.

• Writes require deliberate authorization. A write tool needs both an environment- level flag and a per-call confirmation, and the underlying HTTP client surfaces an ambiguous write outcome for operator review rather than retrying automatically.

• Every error is actionable. Error details are sanitized before they reach the caller, and every error names what to do next.

• Patterns have production provenance. Every pattern in this template is ported from a pattern already shipped in a working server — see the doc comments for provenance.

License

MIT.