iris-eval/mcp-server

This is an early-stage project (24 days old) with minimal adoption signals (6 stars, 2 forks, zero velocity). It positions itself as 'the agent eval standard for MCP,' but lacks the community validation, dataset, or reference implementations to establish that standard. The core functionality—evaluation scoring, safety checks, cost tracking—are well-established patterns in LLM ops tooling (Weights & Biases, Langsmith, custom eval frameworks). The MCP-specific angle is real but niche; MCP adoption itself is nascent. No evidence of users, integrations, or ecosystem gravity. This is a working prototype addressing a real pain point (agent evaluation is fragmented), but it's easily replicated by frontier labs as a feature within their agent platforms (OpenAI's evals, Anthropic's custom evaluators, Google's evaluator pipelines). Frontier risk is high because: (1) evaluation is core to their product roadmaps, (2) MCP is Anthropic-controlled, giving them native leverage, (3) the technical bar is moderate—well-understood metrics and logging. A frontier lab could ship MCP-native eval tooling as a direct feature. The project would need significantly deeper traction (100+ stars, real-world dataset, strong MCP community endorsement) to move into defensibility range 5+. Current score reflects: new repo, tutorial/demo maturity, standard patterns, trivial to clone for anyone with agent eval experience.