On the npm registry, a small package called agnost ticks over to version 0.1.10, last published about a month ago, with nine other projects already pulling it in [npm]. It is a thin SDK with an outsized ambition: give the people building Model Context Protocol servers and AI agents the same kind of product analytics that web teams have taken for granted since Mixpanel.
That is the bet behind Agnost AI, a San Francisco pre-seed founded in 2024 by Shubham Palriwala and Shubham Patel, and backed by Entrepreneurs First and Transpose Platform. The pitch on the homepage is direct: "Product analytics for AI agents. See which intents convert, where users drop off, and what's driving retention" [Agnost AI]. In a category where most teams still grade their own agents by eyeballing transcripts in a Slack channel, that is a useful sentence.
The bet
Agnost sells a conversation-level analytics layer for AI-native products, with a Python SDK for instrumenting agent interactions and native hooks into MCP servers and Claude Code plugins [Agnost AI Docs]. The company tracks tool invocations, task execution, performance, and what it calls satisfaction signals, the small linguistic and behavioral cues that suggest a user got what they came for [Agnost AI Blog]. On top of that telemetry sits an Agent Experience Score, a single composite number meant to function the way a North Star metric does for a consumer app [Agnost AI Blog].
The wedge is MCP. The protocol, originally pushed by Anthropic, has become a default way for agents to call external tools, and operators of those servers currently have very little visibility into what models actually do once a tool is exposed. Agnost's blog leans hard into that gap, with technical posts on long-running task patterns under SEP-1686, plugin usage tracking, and how to lift adoption of an MCP server [Agnost AI Blog]. The X account, joined October 2024, describes the product simply as "Usage Analytics for your MCP Server" [X, October 2024].
Why it could be big
Product analytics has historically been a category that compounds. Once an instrumentation SDK is wired into a codebase, it tends to stay there, and the vendor that owns the event schema tends to own adjacent workflows: dashboards, alerting, experimentation, eventually billing attribution. Agnost is trying to claim that position for a generation of products where the unit of interaction is a conversation rather than a pageview.
The investor signal is consistent with that thesis. Entrepreneurs First is a talent investor that typically writes the first check into technical founders before there is much product, and Transpose Platform has backed early infrastructure bets in the AI tooling stack. Neither firm is a tourist in deep-tech pre-seed. The bet they are underwriting is that conversation analytics for agents becomes its own category, distinct from LLM observability tools focused on prompts, traces, and evaluations.
The early benchmarks Agnost has published hint at where the company wants to plant a flag. In a post defining the Agent Experience Score, the team writes that customer support agents, the most mature category, see good products cluster between 72 and 80, with intent-routed deployments in the top quartile [Agnost AI Blog]. Whether or not that scoring methodology becomes an industry standard, publishing numbered benchmarks is how a young analytics vendor starts to set the vocabulary.
Good support agent score (low) | 72 | score
Good support agent score (high) | 80 | score
Max possible score | 100 | score
The team and traction
Palriwala is cofounder and CEO, confirmed in a recent podcast appearance discussing MCP server metrics [YouTube]. He holds a Bachelor of Technology in Computer Science from Vellore Institute of Technology, graduating in 2023, with prior engineering stints at Onboarding.club, Formbricks, Cisco, and Google Season of Docs [RocketReach]. His public profile describes roughly five years of backend work on secure systems [GetProg.ai]. Patel is the second cofounder [LinkedIn].
Public traction signals are still modest and consistent with a pre-seed timeline: an SDK shipping monthly point releases, nine downstream packages on npm, a steady cadence of long-form technical blog posts, and two recorded podcast conversations including a joint Q&A with Shinzo Labs' Austin Born on MCP analytics [YouTube]. The product is real, the documentation is live, and the founders are publishing into the protocol communities they are trying to sell into.
The honest counterfactual
The bear case is the crowd. LLM observability is already contested by venues like LangSmith, Langfuse, Arize, and Helicone, and general product analytics incumbents are extending into AI use cases. A new entrant has to convince buyers that conversation-level analytics for agents is a separate purchase from prompt-level tracing, not a feature of it. Agnost's answer, visible in the product's framing and the blog's editorial line, is that MCP and agent workflows produce a different event shape: tools, intents, multi-turn satisfaction, plugin usage, long-running tasks [Agnost AI Blog]. If that shape really is different enough to need its own schema, owning it early is valuable. If buyers decide a single trace-and-eval tool is enough, the wedge narrows.
What to watch
The next twelve months should answer two questions. First, does Agnost convert SDK installs into paying design partners with named logos, the kind of reference customers that turn an EF pre-seed into a credible seed round. Second, does the Agent Experience Score get cited by anyone other than Agnost, the leading indicator that a vendor is starting to set category vocabulary rather than just publish into it. A seed round, likely in 2025, would be the natural moment to test both.
If you are building an agent today and your only dashboard is a transcript viewer, which number would actually change how you ship next week?