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 twelve MCP tools

• get_vertical_readiness - the verdict for a business vertical, its four score components, the stack slot blocking it, and three demand fields (paying_businesses, funded_companies, complaint_volume). 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, verified tier, data domain, capability cluster and action class. Every result carries a view_url.
• get_connector - one connector in full, with its tools grouped by verb class, an AI-enriched profile (what it does, data domain, action class, target segment, buildability), its capability cluster, cross-corpus vendor_signal, and a view_url.
• 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, plus the same demand fields (paying_businesses, funded_companies, complaint_volume). 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.
• get_connector_growth - monthly count of new AI connectors listed on the Claude directory, overall or for one category. ChatGPT publishes no listing date, so there is no ChatGPT series. The latest month is partial when it is the current month (see partial_latest_month). A month flagged suspect_gap is a likely source gap, not a real drop. Every result has a view_url.
• get_platform_coverage - which vendors ship an AI connector to Claude, to ChatGPT, or to both, matched on vendor domain. Returns the split plus a list of vendors. Use presence=claude_only or chatgpt_only to find vendors missing from one directory, a distribution gap.
• search_connector_gap - companies from our Funded DB, Stripe Index or TrustMRR datasets that have shipped no AI connector to ChatGPT or Claude. This is a build or partnership list. Matched on exact domain, so absence is a ceiling. Each row links to the company's own page.
• get_action_gap_matrix - every connector tool classified by verb (read, write, transactional, admin, meta, unknown (unclassified)) per data domain. is_action_gap marks domains full of read tools and nearly empty of write tools, where agents can see but not act. Tool counts are mostly Claude, because few ChatGPT apps publish tool lists.
• search_agent_clusters - AI connectors grouped into a fixed set of capability clusters (CRM, payments, calendars...), with member count, nearby complaint demand and a whitespace score (demand per connector). High whitespace = many people describe the problem, few connectors address it.

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.