Armature Launches Product Analytics for MCP Agent Sessions

Founders Theodore and Louis have launched Armature, a product analytics platform designed specifically for the Model Context Protocol (MCP). The tool allows developers to reconstruct full agent sessions—including user intent and agent reasoning—by wrapping their MCP servers with a lightweight SDK.

Armature transforms raw tool-call logs into dashboards that identify popular use cases, frequent failures, and session success rates. To facilitate integration, the company provides SDKs in TypeScript, Python, and Go, which can be implemented in MCP servers, Claude Connectors, and ChatGPT App backends. The platform surfaces user intent, agent thinking, and tool call sequences that are typically hidden inside AI clients like ChatGPT or Claude. To protect data, the tool implements client-side PII and secret redaction before any information reaches Armature's servers. Pricing is free for the first 1,000 sessions per month, with a cost of $50 per 1,000 sessions thereafter.

The Gap in Agent Observability

As more companies expose their products to AI agents via MCP, developers have struggled with a lack of visibility. Traditional analytics platforms, such as Amplitude or PostHog, are built to track human UI interactions. Conversely, LLM observability tools like LangSmith focus on agents built and controlled by the developer. Armature targets the gap between these two, providing "Agent Experience" (AX) analytics for third-party agents interacting with a company's server. This shift is critical as the fundamental assumption that a human is the primary user of software begins to break down, a trend noted by Y Combinator.

Identifying Shadow Use Cases

This visibility allows product teams to discover "shadow" use cases—the unexpected ways users employ agents to interact with a product. More importantly, it helps identify critical failures, such as logic loops or hallucinations, that would otherwise remain invisible. Armature includes automated use-case clustering and issue identification, which can flag problems even when API responses return a 200 OK status. According to data shared by the founders on Hacker News, the instrumentation has a negligible impact on performance, showing an 89.17% pass rate compared to 89.15% over 870 runs.

The Path to Agent Evals

By closing the loop with evaluations, Armature enables a cycle of identifying issues, recommending fixes, and testing them across multiple models before deployment. This process reduces the risk of catastrophic agent errors in production environments. As the ecosystem of MCP-compatible tools grows, the ability to treat agent interactions as a measurable product experience will likely become a standard requirement for developers building agent-facing infrastructure.