Agent Index

What is in the dataset

• 2,800+ MCP servers from the Claude connector directory, with tool lists, auth posture, permissions and declared sensitive data types.
• 4,200+ apps from the ChatGPT directory, with developer, category, version, skills and, where declared, the underlying MCP endpoint.
• 68,000+ individual tool declarations, stripped of vendor prefixes and classified into a verb, an object and an action class.

The public overview is on the Agent Index landing page. The app itself is at /agent-index and requires Pro.

The findings that shape it

• Over 90% of distinct tool names appear on exactly one server. There is no shared vocabulary, so agents cannot route across connectors and every integration stays bespoke.
• Roughly half of all connectors declare no write-capable tool at all. They can read a system but take no action in it.
• The Claude directory grew roughly 10x over 2026, so category saturation moves fast enough that a stale read is misleading.

Agent Readiness Score

One number from 0 to 100 per category, composited from write-capability coverage, connector density, authentication friction and whether the category is carried on both platforms or only one. It answers whether an agent can operate end to end in a vertical or only observe it.

A low score in a category with demonstrated demand is the most interesting signal in the dataset: the market wants it and the plumbing does not exist yet.

The seven MCP tools

• get_vertical_readiness - the verdict for a business vertical, its four score components, and the stack slot blocking it. Lead with the verdict: the numeric score clusters into bands and is not a continuous scale.
• search_agent_ideas - product ideas scoped to a vertical and its blocking slot. Defaults to survivors of an independent adversarial pass.
• search_connectors - search both directories by platform, category, auth posture and verified tier.
• get_connector - one connector in full, with its tools grouped by verb class.
• get_agent_coverage - per-category coverage: how many connectors, what share can act, what share the vendor built.
• get_vertical_stack - the rows underneath a readiness score: every slot the vertical runs on, in order, with vendor counts, agent-surface counts and the blocking slot flagged. get_vertical_readiness returns slots_covered as a count; this returns the stack.
• search_software_vendors - the G2 and Capterra vendors mapped to each slot, filtered by whether they have an agent surface. has_agent_surface: false is the build list.

How ideas are generated

• Capability Whitespace: derived from the connector corpus alone, and must cite a cluster plus a specific empty cell in the gap matrix or a hole in cross-platform coverage.
• Validated Agentic Wedge: promoted only when the gap is corroborated by real complaints from two or more independent sources, with every citation stored and traceable.

Candidates are deduplicated by meaning and then run through an adversarial pass that removes anything whose only moat is shipping an MCP server.