Shriram Vasudevan

Load and latency harness for Model Context Protocol (MCP) servers. Drives real tools/call traffic at configurable concurrency, measures latency percentiles and throughput, and fails CI when a server misses its performance budget. Stdlib-first, zero required dependencies for stdio.

Scaffold a lean, best-practice MCP server that passes the quality checks by construction, with tests, CI and docs built in.

Measure and grade (A to F) the context-token cost of an MCP server's tool surface, with CI gates to stop it bloating.

A zero-dependency, configurable mock MCP server for testing MCP clients and harnesses. Serves a correct MCP handshake over stdio (initialize, tools/list, tools/call). You describe tools and canned responses in a JSON file. Simulates errors and missing-argument handling. Perfect target for your existing conformance, audit, and load-test tools.

Generate human-readable Markdown or HTML docs for any MCP server's tools.

A reference MCP server demonstrating correct use of all three primitives: tools, resources, and prompts. Companion to the AAIF blog "The Unknowns of MCP

mcp-conformance-kit: Making MCP Server Correctness Measurable

A CI-friendly open-source CLI that statically audits MCP servers before they are wired into AI agents. It flags tool poisoning, hidden Unicode payloads, misleading annotations, and lax schemas, without ever invoking a tool, and reports findings by severity for pipeline gating.

a zero-dependency CLI that snapshots an MCP server's tool surface (names, schemas, descriptions, annotations) and diffs two snapshots to catch breaking changes and "rug pulls."

An MCP server that exposes OPC-UA industrial data sources as safe, permissioned tools for AI agents.

mcp-fuzz is a stdlib-only Python robustness harness for AAIF that throws 17 malformed and adversarial JSON-RPC cases at an MCP server over stdio and grades it A to F on crash resistance, responsiveness, error correctness, and envelope hygiene, with a liveness probe after every case so any crash or hang is caught.

mcp-replay records a real MCP session once, then replays it against your next server build to catch behaviour that changed even when the tool schema did not.

An MCP server can quietly rewrite a tool's description or widen its schema after you approve it. mcp-drift-detector pins a trusted baseline and flags the moment that surface moves, separating real rug-pulls from benign changes. Open source, MIT licensed.

A zero-dependency CLI that validates and grades an MCP server's resources and prompts primitives A to F over stdio, filling the gap left by the tool-only checkers and slotting in as a ninth mcp-scorecard dimension

A policy-enforcing runtime proxy for MCP servers over stdio: allow/deny lists, rate limits, argument guards, and audit-only secret redaction. The runtime-enforcement layer of the AAIF MCP toolchain, stdlib-only and dependency-free.

Audits MCP client configs (Claude Desktop, Claude Code, Cursor, Windsurf, VS Code, goose) for risky server wiring like unpinned packages, leaked secrets, shell wrappers and unsafe containers, and grades each A-F with CI gates.