Which businesses can run on autopilot?
Every agentic product dies on the same thing: the systems the business runs on have no agent surface. Agent Index measures that substrate per vertical, so you know whether “run your X on autopilot” is buildable before you build it.
Included with BigIdeasDB Pro
- Healthcare n=67
- 24%
- Travel and hospitality n=65
- 29%
- Business operations n=54
- 31%
- Financial services n=250
- 36%
- Education n=58
- 41%
- Commerce n=106
- 43%
- Legal n=60
- 45%
- Data and analytics n=451
- 46%
- Developer tools n=379
- 53%
- Productivity n=1192
- 65%
- Sales and marketing n=243
- 74%
Most of the map is still blank
The directories make the ecosystem look crowded. Measured properly, it is not.
- 7 in 10MCP connectors were not built by the vendor
- Of the 2,800+ connectors in the Claude directory, which is the one that publishes provenance, only 29% come from the company whose product they connect to. The rest are community wrappers: unofficial, unsupported, and one API change from breaking. A category full of those is a category nobody has really claimed.
- 97.7%of funded companies have no agent surface at all
- We hold funding data on 12,900+ companies with a resolvable domain. Exactly 301 of them appear in either directory. The Stripe Index tells the same story: 97.6% of 29,000+ companies taking payments have no connector either. That is not a missing feature, it is a missing connector, and it is the largest uncovered space in the dataset.
- 10xgrowth in the directory across 2026
- Monthly additions went from double digits to the high hundreds. Open space is real but it is closing, so we track coverage per niche over time rather than reporting it once.
Not every gap is worth building into
A connector that lacks a tool today can add it next release, and your product goes with it. So a missing tool is not an opportunity, it is a countdown. We separate the two and only surface the side that lasts.
Closes with one release
- A connector is missing a write tool
- An endpoint exists in the API but not in the connector
- Auth is awkward, or rate limits are tight
- The tool list is shorter than a competitor's
Filtered out before you see them.
Survives the vendor’s roadmap
- No vendor serves the niche at all, official or otherwise
- The workflow spans several vendors, so no single one will own it
- Exposing it would cannibalise the incumbent's own interface
- The buyer exists and pays, but is too small for the incumbent to chase
These are what Agent Index surfaces.
Autopilot Readiness
One score per vertical, answering one question: can an agent run this business end to end, or only watch it? Readiness is a property of the whole stack, not of any single connector. An agent that reaches the CRM but cannot touch the system the business actually runs on has not automated anything.
- 40%
- Stack coverageHow much of the software a business actually runs on has any agent surface at all.
- 30%
- Write capabilityWhether those connectors can act, or only read. An agent that can read but not write cannot run anything.
- 20%
- Built by the vendorA community wrapper is one API change from breaking, so it counts for less than the vendor's own connector.
- 10%
- Setup frictionAuth posture and required configuration across the covered stack.
The substrate is there. Competition is your risk, not feasibility.
Buildable if you build the blocking piece yourself, or wait for someone to.
Nobody can build here yet. Whoever builds the substrate defines the category.
Every vertical also carries its blocking piece: the one part of the stack whose absence costs the most score. That is the thing to build.
Everything behind the score
- Readiness per vertical
- A score and a verdict for every business type, plus the one piece of the stack that is blocking it.
- The stack behind each business
- Which systems an operator in that vertical actually runs on, and which of them an agent can reach.
- Who is absent entirely
- Companies with revenue, funding and customers that have shipped no agent surface at all.
- Official or community
- Whether a connector comes from the vendor or from a stranger, because only one of those is durable.
- Demand alongside supply
- Paying businesses and complaint volume per vertical, so an empty space is distinguishable from a dead one.
- Ideas with citations
- Scoped to a vertical and its blocking gap, and the premium class carries corroborating complaints.
Ideas that have to prove themselves
An idea that cannot point at a measurable gap does not ship. One whose whole premise is wrapping a model around an existing connector gets removed before you see it.
Capability whitespace
Built from the connector corpus alone. Each one cites a cluster and a specific empty cell in the matrix. Fast to scan, grounded in structure.
- Cites a cluster
- Cites a gap cell
- Names the bounding connectors
Validated agentic wedge
Promoted only when the gap is corroborated by real complaints from two or more independent sources. Every citation is stored, so you can follow any claim back to the words somebody actually wrote.
- Two or more independent sources
- Every claim traceable to a quote
- Survived a pass whose only job is removing ideas
Query it from your own agent
Agent Index is not only a page you read. The whole corpus is on the BigIdeasDB MCP server as seven tools, so Claude, ChatGPT, Cursor, VS Code or any MCP client can search the connectors and pull a vertical’s readiness mid-conversation. Ask “can an agent run a dental practice yet, and what is missing?” and the answer comes back from this dataset rather than from the model’s memory.
Server URL
https://bigideasdb.com/api/mcp/bigideasdb-mcpGenerate a client ID and secret in the dashboard, add the URL to your client, and the Agent Index tools appear alongside the other 30. MCP access is part of Pro.
- get_vertical_readiness
- Autopilot Readiness for a business vertical: the verdict, the four score components, and the stack slot blocking it.
- search_agent_ideas
- Product ideas scoped to a vertical and the slot an agent cannot reach, defaulting to the ones that survived the kill pass.
- search_connectors
- Search both directories at once by platform, category, auth posture, verified tier and whether the connector can act.
- get_connector
- One connector in full: vendor, auth posture, endpoint, permissions, and every tool it declares grouped by verb class.
- get_agent_coverage
- Per-category coverage: how many connectors serve it, what share can act, what share the vendor built. Weakest first.
- get_vertical_stack
- The rows underneath the verdict: every slot the business runs on, in order, with how many vendors fill it and how many an agent can reach.
- search_software_vendors
- The 22,000+ vendors mapped to those slots. Ask for the ones with no agent surface and you have the build list.
Same data as the app, same rounding and the same caveats. can_act comes back null when a connector publishes no capability list, which is not the same as being unable to act.
The research behind it
Three deep dives published from the same corpus, with the method and the caveats written out in full.
Frequently asked
- Agent Index scores business verticals on whether AI agents can actually run them. For each vertical it maps the software an operator in that business genuinely relies on, checks which of those systems have an agent surface and whether that surface can act or only read, and produces an Autopilot Readiness score plus the single missing piece blocking the vertical. It is built on a census of 7,000+ connectors across the ChatGPT apps directory and the Claude MCP directory.
- Because the thing you are building is a product for a business, not a connector. Integration is the documented bottleneck for vertical AI agents, so the question that decides whether your product is possible is whether the systems that business runs on can be reached and written to. That is a property of the whole stack. A directory listing cannot answer it.
- Four inputs, weighted: stack coverage at 40% (how much of the software the business runs on has any agent surface), write capability at 30% (whether those connectors can act or only read), vendor-built at 20% (a community wrapper is one API change from breaking, so it counts for less), and setup friction at 10%. Scores are banded rather than read as exact numbers, because the inputs carry real uncertainty and a band survives that where false precision would not.
- Under a third of connectors in the directory were built by the company whose product they connect to. The rest are community wrappers: unofficial, unsupported, and liable to break when the underlying API changes. If you build a product on a community wrapper, you have taken a dependency on a stranger's weekend project. The score treats vendor-built and community-built differently for that reason.
- Yes, and that is exactly why we filter those out. A connector that lacks a write tool today can add it next release, and any product built on that gap goes with it. We separate gaps that close with one release from gaps that are structural: nobody serves the niche at all, the workflow spans several vendors so no single one will own it, exposing it would cannibalise the incumbent's own interface, or the buyer is too small for the incumbent to chase. Only the second kind is surfaced.
- Two classes, and you always see which is which. Capability Whitespace ideas come from the connector corpus alone and must cite a vertical and a specific blocking gap. Validated Agentic Wedge ideas are only promoted when that gap is corroborated by real complaints from two or more independent sources, with every citation stored so you can trace any claim back to what someone actually wrote. Anything whose only moat is shipping an MCP server is removed before you see it.
- That connection is what makes the scores possible. Demand per vertical comes from businesses already taking payments in that niche and from the 1M+ complaint corpus. We can also identify companies with revenue, funding and customers that have shipped no agent surface at all, which is both a gap analysis and a build list. No directory can produce that.
- Yes. Agent Index is on the BigIdeasDB MCP server as seven tools, so any MCP client can query the corpus directly instead of reading the page. get_vertical_readiness returns the verdict and blocking stack slot for a business vertical. get_vertical_stack returns the slot rows underneath that verdict, with vendor and agent-surface counts per slot. search_software_vendors returns the G2 and Capterra vendors mapped to those slots, including the ones nothing connects to. search_agent_ideas returns ideas scoped to a vertical and its gap, with the stored citations when you ask for them. search_connectors searches both directories by platform, category, auth posture and whether the connector can act, get_connector returns one connector with its tools grouped by verb class, and get_agent_coverage returns per-category coverage weakest first. Add https://bigideasdb.com/api/mcp/bigideasdb-mcp to your client with the client ID and secret generated in the dashboard. MCP access is part of Pro.
- Generate a client ID and secret on the MCP page in your dashboard, add the server URL https://bigideasdb.com/api/mcp/bigideasdb-mcp to your client with X-Client-Id and X-Client-Secret headers, then restart the client. The seven Agent Index tools appear alongside the other 30 BigIdeasDB tools. It works with Claude, Claude Code, ChatGPT, Cursor, VS Code, Windsurf, JetBrains and any other MCP-compatible client, because the server is one hosted HTTP endpoint rather than a per-client integration.
- The readiness score and verdict for each vertical are free. The stack detail, the blocking gap, the demand evidence and the ideas are part of the BigIdeasDB Pro plan, alongside Stripe Index, Revenue Intelligence, Funded DB, SellSide DB, and the rest of the suite.
Build where nobody has
7,000+ connectors mapped, so you can find the niche that has none.