MCAFax Lab Journal, November 5, 2026
Abstract
Four popular open-source "AI second brain" projects all store memory as plain Markdown in an Obsidian vault and drive it with agent skills. We read each repo and asked which parts make sense when the second brain is a public website and not a private vault. Our answer: keep the file conventions (short skills, cited ingest, a linter, a shared log) and drop the vault itself. MCAFax.com already does the vault's job. The Notebook is where things get captured. The Lab Journal is where they get consolidated.
Methods
- Source: a post on X listing the four repos (https://x.com/lez99999/status/2118063491894714842, 2026-11-04 UTC).
- On 2026-11-05 we pulled each repo's metadata (stars, license, last push), its file tree, and its README from the public GitHub API. We did not run any installers.
- We sorted every component into three kinds: skill (a prompt file), code (runs with no model calls), or template (a vault scaffold). Then we matched each one to an MCAFax surface.
Data (as fetched 2026-11-05)
| Repo | Stars | License | Shape |
|---|---|---|---|
| kepano/obsidian-skills | 50,807 | MIT | 6 SKILL.md files + references/ |
| AgriciDaniel/claude-obsidian | 16,190 | MIT | 15 skills, 3 agents, rule files for multiple agents, checksums |
| eugeniughelbur/obsidian-second-brain | 5,168 | MIT | 48 commands, 8 agents, CLI adapters, "Crew" logbook |
| coleam00/second-brain-starter | 870 | none stated | 1 PRD-generator skill + memory templates |
Results
- Short skill, heavy references. kepano's skills follow the Agent Skills spec (agentskills.io). Each SKILL.md stays small, and the detail lives in
references/files that the agent reads only when it needs them. That matches our rule that a skill is just a trigger. - Spend fewer tokens on ingest. kepano ships a
defuddleskill that strips web pages down to clean Markdown. The README gives the reason as "removing clutter to save tokens." - Cite, then lint. claude-obsidian keeps original sources untouched in
inbox/so every summary can link back to its source. It also ships dedicatedwiki-lintandverifieragents. - Humans decide what becomes memory. In obsidian-second-brain's Crew, agents share one plain-folder logbook. Per its README, a lesson becomes a rule only when a human runs
keep, and Crew runs no background loop. - The cost of a big command surface. The 48 commands and 15 skills give agents more triggers to choose between. Our minimal-token approach goes the other way.
- Licensing. second-brain-starter has no license file, so we can learn from its intake to PRD idea but cannot copy its text or code.
Discussion
A private vault makes memory visible only to its owner. When the website is the vault, every write-up does double duty: it is memory for our bots and it is public proof of work. The Obsidian-specific parts (Bases, Canvas, the graph view, the CLI) don't apply to us. The patterns underneath them do: plain files, a cited ingest step, deterministic linting, and human-gated rules. These carry over directly and fit our brain model. (1) Code does the work: fetch, clean, lint, render. (2) A one-line skill triggers it. (3) A Lab entry explains how it works.
Limits. We counted stars at a single point in time. We did not benchmark token savings. We did not audit the installers.
Next builds (ranked by token efficiency and cutting-edge fit)
lab_lintin site_audit: pure code that checks Notebook/Lab for missing citations, broken links, orphan posts, and missing scrub markers.clipingest: a Defuddle-style step (via the open-source library, no SaaS) that turns a URL into clean Markdown plus a source stub, as the first step of every research run. Triggered by a one-line skill.- Public skill catalog page: generated by code from our SKILL.md frontmatter, Agent Skills spec-compliant, in the style of knap's template + data to page.
- Crew-style bot logbook: an append-only JSONL file shared by all bots. A human
keeppromotes a lesson to a rule. The weekly org-audit reads it. - AGENTS.md to site rules mirror: one source file rendered into each agent's rule file and into a public "how we work" page.
- Weekly review skill: reads the past week's Notebook entries and drafts Lab Journal candidates. Run on demand, not by polling.
- Apply-funnel intake to plan generator, an original implementation of second-brain-starter's form to PRD idea.
Sources
- https://github.com/kepano/obsidian-skills
- https://github.com/AgriciDaniel/claude-obsidian
- https://github.com/eugeniughelbur/obsidian-second-brain
- https://github.com/coleam00/second-brain-starter
- https://agentskills.io/specification
- https://x.com/lez99999/status/2118063491894714842
Appendix: how each repo maps to MCAFax.com
- kepano/obsidian-skills: keep the short-skill pattern with on-demand references; adapt defuddle as our ingest step and knap as zero-AI page generation; skip Bases, Canvas and the CLI.
- AgriciDaniel/claude-obsidian: keep the ingest, cite and lint loop; adapt wiki-lint as deterministic code and one AGENTS.md mirrored to every agent; skip the vault and graph UI.
- eugeniughelbur/obsidian-second-brain: keep the shared logbook with human-gated rules; skip the 48-command surface and the curl-pipe installer.
- coleam00/second-brain-starter: adapt the intake-to-plan idea only, since there is no license; skip 30-minute heartbeat polling.
Rule of thumb: keep plain files, deterministic code and one-line triggers; skip anything that needs Obsidian, SaaS or recurring model polling.
