# Social media share of new MCP repositories on GitHub: method and gate results

Growth test added on 13 October 2026 to Blotato's study MCP Servers for Social Media Automation: 2026 Statistics, https://www.blotato.com/research/social-media-mcp-servers#growth. Gate figures come from gates.json in the same folder. Completeness counts come from the raw search results, and spot-check counts from the label-assembly step, and neither is in the downloads.

## Gate results

Every gate had to pass. A low-end figure multiplies the observed change by the low end of a 95% range from the labeled samples: for G1, the change in the share of sampled social media matches that are real servers, and for G4, that change divided by the change in the share of sampled MCP repositories that contain MCP code.

| Gate | Question | Result | What we found | Bar |
|---|---|---|---|---|
| G0 | Enough social media matches to test | Pass | 1,223 keyword matches in the first period (owner-capped, not servers) | at least 30 |
| G1 | More social media matches in the later period | Pass | 2.35x, low-end 2.26x | low-end 1.25x or more |
| G2 | Match counts not past a peak | Pass | Jul to Sep 2026 was the highest quarter | latest quarter's count at or above the one before, and at least 80% of the earlier peak |
| G3 | More search and download demand | Pass | US searches 4.48x (6 terms, Mar to Aug 2026 against Sep 2025 to Feb 2026), npm downloads 1.21x (164 packages, the study periods) | searches 1.10x or downloads 1.25x, each with its largest term or package removed, and searches not below 0.80x |
| G4 | Social media share of new MCP repositories at least 10% higher in the later period | Fail | 1.28% to 1.57%, low-end increase 7.2%, latest quarter 1.60% against 1.53% | low-end increase 10% or more, and the latest quarter's share at least 90% of the one before |
| G5 | G1 and G4 hold without matches whose only match term is the largest group (YouTube), and without the top month (August 2026) | Fail | G4 low-end increase without YouTube-only matches 9.8%, without August 2026 5.0% | G1 and G4 still pass |

## Method

**What was compared.** GitHub repositories created from October 2025 to March 2026 against those created from April to September 2026, forks excluded, found through GitHub's search in time windows of under 1,000 results each. A social media match names one of the 10 platforms on this page, or social media in general, plus MCP. Matching to the MCP list is by repository, and the share counts only matches that also appear in it. Then each owner counts once a month in each list.

**How complete the lists are.** The social media list matched GitHub's own counts exactly. For the MCP list, GitHub's counts are 104 repositories higher than what its search returned for the later period and 8 higher for the earlier one. A window short by up to 0.5% of its count, or by 2 for small windows, counted as complete, and 54 of the later period's 104 sit in the 14 windows that stayed 3 to 5 short after a second pull. Because the social media list is complete, the only social media matches the MCP list can be missing are the 26 earlier and 21 later repositories (counted before the owner cap) that GitHub's MCP search did not return, which the share leaves out. Any other missing repository has no social media match. Adding every one of the later ones to the MCP list, none of the earlier ones, and the earlier period's 8 to its MCP list would lift the low-end increase to 7.9%, still under the bar.

**How matches were checked.** We drew 150 social media matches per period at random, split by match type (each single platform, two or more platforms, and social media with no platform named) in proportion to their numbers, and 100 MCP repositories per period. In both lists an owner's earliest-created repository that month stood in for them. Because the split is proportional, the samples need no weighting. AI models labeled each one from its saved README and manifest under written rules. A second AI model independently re-labeled a seeded 100 of the 500. It differed on 1 of the yes-or-no labels the gates use and on 1 label for whether a server publishes, and under our written rule its label replaced the first one on those 2. That shows agreement between two models, not accuracy against a person reading the files. In the samples, 32% of earlier matches and 41% of later ones were MCP servers acting on a social platform's accounts or content, and 67% and 70% of MCP repositories contained MCP code.

**The two estimates.** The share gate takes the observed ratio of the two shares (1.23x) and multiplies it by the low end of a 95% range for the change in the social media real-server rate divided by the change in the MCP-code rate, taken as the 2.5th percentile of 10,000 resamples, each sample resampled as a whole as the registered code does, not within its match types. That gives a 7.2% increase. The range covers sampling only, not errors the two AI models share or repositories GitHub's search did not return.

**Robustness.** Both checks failed on the share gate and passed on the growth-in-matches gate (G1). Leaving out matches whose only match term is YouTube, the largest group, from the social media side and its sample, the low-end increase in share is 9.8%, while G1 is 2.55x with a low end of 2.29x. Leaving out August 2026, the month with both the highest share and the highest count, from both lists, the low-end increase in share is 5.0% and G1's low end is 2.19x on a per-month basis.

**Demand.** US Google searches for 6 social media MCP terms grew 4.48x, March to August 2026 against September 2025 to February 2026, the 12 months our data source returned, with the largest of 7 terms removed. npm downloads of 164 packages tagged mcp and a social media term, published before October 2025, grew 1.21x from October 2025 to March 2026 to April to September 2026, with the largest of 165 removed. Searches and downloads are not users.

**When the rules were set.** We registered the test on 12 October 2026, before collecting any data, with the 10% bar and the plan to publish only if it passed. About six minutes after collection began, before any figure was computed, we revised it. The revision made the share gate stricter by also adjusting for the share of MCP repositories that contain MCP code and adding the latest-quarter check, raised the social media samples from at least 100 to 150 per period and added the MCP sample, redefined the robustness checks, loosened the completeness rule from an exact match to the tolerance above, and set a minimum for the search data. With the earlier share formula, applied to these same samples, the share gate's 95% low end would have been 1.17x (16.7%) against its 1.10x bar. We report the revised one.

**Changes after that.** Each was logged before any growth figure was computed. The share counts only matches also in the MCP list, which slightly raised the observed change and its low end. The rule for X matches was copied exactly from the count above, and one robustness definition was matched to the code. An npm search that did not filter by platform was replaced with a tag search. A script that re-pulled short windows was fixed and its first run discarded. A rule for settling the two models' disagreements was fixed before any sample was drawn. Last, the completeness rule stopped the study once collection finished, and, counting whole windows as the locked code did, so would the earlier exact-match rule have. We changed it to count missing repositories instead of whole windows, and the study continued.

**Publishing.** We decided to publish the result anyway on 13 October 2026, after seeing it. No gate, threshold, sample or label was changed after the result.

**Limitations.** We read each repository's current name, description and topics, so a repository that added MCP or social media words later is counted in the month it was created. Repositories deleted before we checked are missing. Hosted servers with no public repository, Blotato's included, and servers outside GitHub are not in these figures. A match is a keyword match, and only the samples were checked by reading the files. The registration and its timestamps are in our own change log, and no outside service time-stamped them. The repository lists and the saved files the models read are not in the downloads, only the quote behind each label. The scripts collect the lists again from GitHub, whose results change over time.