mcpup

Semantic health monitoring for MCP servers. A "Checkly for MCP" that catches what generic uptime monitors miss: silent failures, schema drift, broken tools.

See mcpup-brief.md for the full product context.

Packages

Package	Role
@mcpup/core	Probe engine: initialize handshake, tools/list, strict parsing, schema snapshot & drift. Reusable by the CLI and the future SaaS.
mcp-check	Open-source CLI: one-shot semantic health check. Top-of-funnel of the product.

Development

Bun monorepo (runtime, tests, bundler).

bun install          # install workspace dependencies
bun test             # runs the whole suite (core + CLI e2e)
bun run check <url>  # run the CLI against an endpoint

# per package
cd packages/core && bun run typecheck
cd packages/cli  && bun run typecheck

Status

Building block 1 of the brief (§10): the probe client that de-risks everything else, shipped as the mcp-check CLI.

• ✅ initialize handshake + tools/list over Streamable HTTP transport (official @modelcontextprotocol/sdk).
• ✅ Silent failure detection (200 OK hiding a JSON-RPC error / unreadable payload).
• ✅ Schema snapshot + drift detection (structural diff vs baseline).
• ✅ Safe tool call (--call): args synthesized from the schema, validation of the shape of the response (isError, outputSchema mismatch via Ajv).
• ✅ mcp-check CLI with CI exit codes (0/1/2/3/4) and --json output.
• ✅ Validated by tests against a hand-rolled fixture and a real MCP server from the SDK.

Next building blocks (out of current scope)

Scheduling & history · alerting (email/Slack/webhook) · dashboard · status page.