---
title: "agent-seer"
description: "Specification-driven evaluation, deterministic capability linting, and synthetic test harness generation for Model Context Protocol (MCP) tool suites."
canonical: https://agentpluginsdirectory.com/plugins/agent-seer
last-updated: 2026-10-21
---

# agent-seer
Specification-driven evaluation, deterministic capability linting, and synthetic test harness generation for Model Context Protocol (MCP) tool suites.
- Slug: agent-seer
- Publisher: ghchinoy
- Repository: https://github.com/ghchinoy/ai-paper-reproductions
- Manifest: agent-seer-mcp-tool-calling/plugin.json
- Version: 1.0.0
- License: Apache-2.0
- Category (editorial): agent-tooling
- Skills: 1 (agent-seer)
- MCP servers: 0
- Stars: 1
- Repository created: 2026-08-30
- Repository last pushed: 2026-08-31
- Publisher type: User
- Listing: https://agentpluginsdirectory.com/plugins/agent-seer
- Schema: https://agent-plugins.org/schemas/1.0.0/plugin.schema.json

## What agent-seer does, in the publisher's words

A spike report on spec-driven tool-calling evaluation, the negative result we hit, and the one change that fixed it.

We ship a set of Model Context Protocol (MCP) servers that let an agent drive Google Cloud's generative-media APIs: text-to-video, image-to-video, text-to-speech, music, and audio/video compositing. Our existing checks confirm that a server starts, answers a tools/list handshake, and produces a non-empty artifact. They do not confirm that an agent calls those tools correctly: right tool, right order, schema-legal arguments. That orchestration layer is where real bugs reach users. An agent picks a model that cannot do what the prompt asks, passes an aspect ratio the model rejects, or breaks an image-to-video chain by forgetting to pass the upstream frame.

A recent paper offered a way to build that missing layer without hand-authoring a benchmark. We reproduced its method against one of our servers to see whether the signal holds up on real generative-media tool schemas.

From the project README, punctuation lightly normalized. Full text: https://raw.githubusercontent.com/ghchinoy/ai-paper-reproductions/HEAD/agent-seer-mcp-tool-calling/README.md

## Skills

- agent-seer: Specification-driven evaluation, deterministic capability linting, and synthetic test harness generation for Model Context Protocol (MCP) tool suites. Validates tool-calling transcripts against schemas and runtime capability matrices in <1ms without LLMs, and scores semantic orchestration using a d…

Descriptions come from the frontmatter of each SKILL.md, punctuation lightly normalized.