We built a test column in our demo environment and planted 47 components of 32 kinds around it, so every kind of component that can read or write a column does. Then four Claude models got the same question in four setups, three runs each. Without Pathix, none of the 36 runs found all 47. With the Pathix™ MCP connected, all 12 did.
Every run got the same prompt: find every component that writes to the Risk Score column on the Submission table and every component that reads it, and say for each one whether it was proved from the component's definition or logic, or only inferred. The models were Claude Haiku 5.5, Sonnet 5.5, Opus 5.5 and Fable 5.1, run headless in Claude Code 2.1.304 at effort high, with every write tool blocked.
The Dataverse MCP's read-only tools, and nothing else.
Adds the Dataverse Web API.
Adds the Web API and the customer's plugin source code, with tools to read and search it.
The first setup again, with the Pathix MCP connected beside the Dataverse MCP.
Each cell is one setup and model's median run of three, by components found: how many of the 47 it found, the claims it got wrong, its time to evidence (from the start to the last tool result it used) and its cost, estimated at Claude API list prices. The bottom line gives the range across all three runs.
Demo environment · a test column built for this benchmark · 2026-10-20 · Claude Code 2.1.304 at effort high · cost estimated at Claude API list prices from each run's tokens; the runs were billed to a Claude plan seat
More access didn't buy more answers. The Jr. Analyst's best was Opus's 42, and with the Web API added, Opus found 37.
Without Pathix, the most-missed kinds were the ones no one thinks to check: Power Pages server logic was missed in 36 of 36 runs, and the dataflow in 32 of 36.
With Pathix, every model found all 47, and Haiku did it in under thirty seconds for about two cents at API list prices.
Out of the 36 runs without Pathix, how many missed each of these. The runs with Pathix found every one of them.
Most columns aren't this busy, but you only know that after looking everywhere, so a column with any history has to be checked against every kind of component. The runs without Pathix looked at between 14 and 32 of the 32 kinds.
Pathix had already read all 32 of these kinds of component when it scanned the environment, so the agents asked Pathix for the answer instead of searching each kind themselves. That scan ran before the benchmark, and its time isn't in any run's time. The Pathix runs made 6 to 47 tool calls each.
The Dataverse MCP was in all four setups, the Pathix one included, because Pathix works alongside it. What makes the question hard is its shape, proving absence across 32 kinds of component, not the tool.
Pathix MCP · read-only · signs in as you · metadata only