First-party observed usage
Which MCP servers are actually called in coding sessions?
This snapshot groups exact tool names shaped like mcp__server__tool into server-level totals. It ranks observed invocation activity—not installs, GitHub stars, registry popularity, satisfaction, or industry-wide market share.
What the data says
Claude In Chrome leads this eligible snapshot by calls with 61,170 observed invocations across 8 exact tool names, with an observed error-call rate of 5.8%.
Eligible server calls
158,955
Across servers meeting the 5-call floor.
Attributed turn tokens
45.69B
Model-turn tokens attributed to MCP calls, not MCP execution cost.
Observed error calls
9,502
6.0% of eligible calls.
Eligible servers
20
0 below the public ranking floor.
Claude In Chrome
8 exact tools · 17.66B attributed turn tokens
calls61,170
What the MCP usage metrics mean
- Calls
- The number of observed invocations reported for exact MCP tool names in participating users' latest transcript snapshots.
- Attributed turn tokens
- Tokens from the model turn containing the tool call, divided across tool calls when a turn invokes more than one. This is not the MCP server's compute, network, or API token cost.
- Error calls
- Calls whose transcript result was observed as an error or interrupt. The server-level value can be summed because each exact tool row counts calls, not distinct people.
- Error-call rate
- Observed error calls divided by observed calls. It is a transcript outcome signal, not an uptime SLA or a controlled reliability test.
Popularity, workload, and reliability answer different questions
Call volume shows how often an MCP server's observed tools were invoked. Attributed tokens reflect the surrounding model work and can be high for context-heavy tasks even when the server itself returns a small response. Neither measure establishes adoption outside this cohort.
Error rate is descriptive, not causal. A failed call can reflect authentication, invalid arguments, network conditions, permissions, upstream services, or the server. Compare it only with the sample size and exact tool mix visible on this page.
Methodology and limitations
Snapshot coverage. The page reads toolCalls from the public insights response. These values come from each listed participant's latest submitted transcript snapshot; they are overwritten on a new submission and are not a trailing-30-day time series. The displayed generation date identifies the API snapshot rendered here.
Server grouping. Only exact names matching mcp__server__tool are included. The text between the first and second double-underscore separators is the server key. Calls, attributed tokens, and errors are additive across exact tool rows; distinct-developer counts are not aggregated because the API does not expose a server-level union.
Sample floor. A server must have at least 5 observed calls to enter the ranking. This reduces one-off noise but does not make the sample representative or a statistical benchmark.
Result ceiling. The API returns at most 100 skill/tool-call rows after ranking built-in and MCP tools together. Lower-volume MCP tools may therefore be absent, so the table is a ranked view of returned observations—not a complete registry of every installed or invoked MCP server.
Cohort and attribution limits. Public participation is opt-in and not weighted to represent all AI coding users. Token values belong to the model turn containing each call and are split across calls in multi-tool turns. Do not interpret them as the MCP server's direct execution cost, latency, or resource use.
Sources
- whoburnedmore public insights API
Aggregated model, agent, trend, skill, and tool-call totals from participating developers.
- Data collection and privacy methodology
Explains what the CLI aggregates, what it does not collect, and the limits of the public cohort.