Lighthouse

Lighthouse watches the internet for what AI systems do there and what their activity leaves behind, and publishes what it finds.

Studies LH002

LH002 cohort

The ten repositories as fixed in brief.md before collection, with what each source returned for each. Every figure here is from a table in data/, read on 26 September 2026; the window is 2026-03-26T00:00:00Z to 2026-04-22T23:59:59Z. No repository was added, dropped or replaced after the brief was written. Contributor identities appear only as counts and classes; the pseudonymous per-author table is data/authors_pseudonymous.csv.

Membership and selection

id repository PyPI package how it entered first commit in git first PyPI release
R00 https://github.com/modelcontextprotocol/python-sdk mcp the library 2024-09-24 2024-11-20
R01 https://github.com/saigontechnology/agentcrew agentcrew-ai frame A, hash order 1 2025-02-18 2025-06-23
R02 https://github.com/robertwhiffin/ai-slide-generator databricks-tellr-app frame A, hash order 2 2025-11-05 2026-01-19
R03 https://github.com/ooo0ooo/lean-lsp-mcp lean-lsp-mcp frame A, hash order 3 2025-03-29 2025-03-31
R04 https://github.com/paulieb89/pyp6xer-mcp pyp6xer-mcp frame A, hash order 4 2026-04-04 2025-12-08
R05 https://github.com/oraios/serena serena-agent frame A, hash order 5 2025-03-23 2025-07-21
R06 https://github.com/eddmpython/dartlab dartlab frame A, hash order 6 2026-03-06 2026-03-06
R07 https://github.com/mistralai/mistral-vibe mistral-vibe frame A, hash order 7 2025-12-09 2025-11-17
R08 https://github.com/drakkar-software/octobot octobot frame B, oldest first release 2018-02-23 2019-01-16
R09 https://github.com/jesse-ai/jesse jesse frame B, second oldest 2020-04-07 2020-04-06

Frame A is the 34 direct dependents of mcp 1.27.0 that Open Source Insights lists; frame B the 95 of mcp 1.26.0. The rule and the counts excluded at each step are in brief.md; the script is scripts/select_cohort.py and its full output, with every excluded package and the reason, is reproduced by running it.

Inclusion notes. Fork status could not be read: GitHub's REST API answered 403 through this session's proxy, and the proxy's rule is to report such a refusal, not route around it. Every repository here is the one its package's own metadata names, so none was chosen as a fork, but none is confirmed not to be one. R04's GitHub history begins on 2026-04-04 with a single commit holding the whole project, and its PyPI release of 2026-03-26 names a GitLab repository as its source (Open Source Insights, UNVERIFIED_METADATA); the nine days before are missing, not zero. R07's first commit in git, on 2025-12-09, is later than its first PyPI release, on 2025-11-17, so its earlier history is not on GitHub either; that falls before the window and does not affect its readings. No paths matched the vendored-content rule. Lockfiles are counted apart from other lines. R06's changed lines are 91 per cent JSON files (data/changed_lines_by_location.csv), which the tables keep visible rather than exclude.

Readings over the window

id observed days active days commits per observed day bot / human account / unknown commits PyPI releases git tags mcp in git at start at end
R00 28 13 95 3.39 2 / 79 / 14 1 0 (the library) (the library)
R01 28 20 97 3.46 0 / 0 / 97 17 17 >=1.24.0, lock 1.25.0 >=1.24.0, lock 1.26.0
R02 28 13 210 7.50 0 / 23 / 187 0 0 none >=1.0.0, lock 1.27.0
R03 28 3 26 0.93 0 / 0 / 26 2 2 ==1.26.0, lock 1.26.0 ==1.27.0, lock 1.27.0
R04 19 5 13 0.68 0 / 0 / 13 1 0 no history ==1.27.0, lock 1.27.0
R05 28 28 435 15.54 9 / 42 / 384 4 4 ==1.26.0, lock 1.26.0 ==1.26.0, lock 1.26.0
R06 28 28 1,482 52.93 1 / 1,481 / 0 35 50 >=1.0, lock 1.26.0 >=1.0, lock 1.27.0
R07 28 8 9 0.32 0 / 0 / 9 9 9 >=1.14.0, lock 1.26.0 >=1.14.0, lock 1.26.0
R08 28 16 124 4.43 1 / 0 / 123 1 1 ==1.26.0 (nested file) ==1.26.0 (nested file)
R09 28 9 17 0.61 0 / 0 / 17 4 0 none none

Sources: data/repo_summary.csv, data/pypi_releases_in_window.csv, data/tags_in_window.csv and data/mcp_state_window_start_end.csv. "Human account" means an author address at GitHub's noreply domain with no bot marker; "unknown" is every other author without a bot marker, which includes people who commit under a personal or work address. R00's release tag is not counted as a window tag because it is a lightweight tag and git gives it the date of its commit, 24 March.

Coverage by source

id git over HTTPS Open Source Insights graph for the window-end version Software Heritage GitHub origin GH Archive, 2 April 14:00 UTC
R00 complete resolved archived, window-end head present 0 events
R01 complete resolved not archived 0 events
R02 complete resolved, no mcp in that version not archived 2 events
R03 complete resolved not archived 0 events
R04 from 4 April resolved not archived 0 events
R05 complete for surviving branches resolved archived, window-end head present 2 events
R06 complete resolved not archived 0 events
R07 complete resolved archived, window-end head present 0 events
R08 complete resolved archived, window-end head present 0 events
R09 complete failed: the service reports a parse error in a requirement marker archived, window-end head present 0 events

"Complete" means every commit reachable in September 2026 from a branch or tag; commits on branches deleted since are not in it (see LH002.md, Findings). Sources: data/software_heritage_coverage.json, data/depsdev_graphs_window_start_end.csv, data/gharchive_hour_vs_git.csv.

From studies/LH002/cohort.md in the repository, last changed 26 September 2026.