COG Second Brain
A self-evolving, local-first second brain that turns daily braindumps, team intelligence, and PM workflows into verified knowledge using plain Markdown files and AI agents.
README explicitly describes least-privilege design (read-only verifiers cannot write files, workers return only status + path), per-file confirmation before updates, and publishing guardrails (volume caps, screenshot-verified posting). But sensitive-data handling is one line ('local md files') while People CRM, meeting transcripts, and Slack/Linear/PostHog sync actually push personal data into the vault and model APIs with no masking or consent mechanism shown; external mutations (publish-to-confluence, two-way Linear sync, worker-executor 'pre-approved mutations') lack approval detail; dependency security has only an npx install hint, no lockfile or audit evidence.
README is internally consistent: 33 skills, 6 workers, 4 verifiers match across sections; integrations are declared to 'degrade gracefully' with SETUP.md pointers. Failure messaging appears only sporadically (fix-agent max 2 retries); no systematic evidence, hence deduction.
Excellent audience/scenario coverage: role packs for PM/eng lead/designer/founder, 'Try saying...' trigger examples for all 33 skills, and a support matrix with per-platform coverage and fallback paths. Environment fit across Claude Code/Cursor/Kiro/Gemini/Codex with a concrete skill-discovery table.
Clear information architecture (numbered 00–06 vault, framework/content separation), complete install notes (clone + onboarding table + skills.sh), practical FAQ, and a full MIT license. Weaknesses: known limitations are barely disclosed ('zero maintenance', '95%+ source accuracy' are unverified marketing claims); versions exist (COG-VERSION, roadmap v3.1–v3.7) but no standalone changelog in this evidence; solo maintainer, no governance statement.
Output usability is evidenced by design (/tmp file + path passing to cut token bloat, post-condition checks, review-cockpit approval slots). Marginal value vs Notion/Obsidian is clearly positioned. Cost/benefit honestly notes users pay their own API costs. But all efficacy claims (120+ braindumps, 5 strategic insights) are self-reported without evidence.
Claim traceability is weak: headline metrics (95% source accuracy, 120+ entries) cite nothing; external corroboration is limited to directory-site listings; facts and inference are blended in marketing prose, with anthropomorphic claims ('self-evolving', 'learns your patterns') outrunning the described mechanics (md files + classification rules). The verification framework (closed-loop, retro auditing evidence quality) is a good idea but is design documentation, not execution evidence.
- People CRM and meeting transcripts push colleagues' personal data into local files and model APIs; the README offers no masking, consent, or GDPR-style handling — assess compliance before team use.
- External mutations (worker-executor, two-way Linear sync-back) claim 'pre-approved' status with thin detail; audit the actual SKILL.md permission gates rather than trusting the README.
- Efficacy claims ('95%+ source accuracy', '120+ braindumps', 'zero maintenance') are self-reported and uncorroborated; do not rely on them when evaluating utility.
- cog-update.sh advertises --dry-run and per-file confirmation, but the implementation is not in the provided evidence; read the script before use.
- npx skills add executes third-party package code from an unverified publisher; review the repository before installing.
What does this agent do, and when should you use it?
COG (Cognition + Obsidian + Git) is an MIT-licensed second-brain framework built entirely on local .md files — no database, no vendor lock-in. It ships 33 AI skills, 10 agents (6 workers plus 4 read-only verifiers), and a people CRM with tiered, evidence-cited profiles. Its execution model is deliberate cost engineering: Sonnet workers handle data-heavy extraction and write results to /tmp/ files, while the lead session reads those files and does the reasoning; verifiers receive paths only, so they observe the artifact itself rather than the worker's summary, and the worker never grades its own homework. An opt-in closed-loop harness walks work through a V-model lifecycle — decompose into falsifiable criteria, build, then verify each criterion with traced evidence. Multi-agent support is a first-class feature: native surfaces for Claude Code, Antigravity, Cursor, Kiro, Gemini CLI, and OpenAI Codex, with AGENTS.md as a universal fallback and an Agent Plugins (spec 1.0.0) manifest. Data syncs across devices via Git and iCloud, and framework updates never touch your personal content.
After cloning and running onboarding in your agent, you can: capture raw thoughts with braindump and automatic domain classification; get daily-brief news intelligence enforcing 7-day source freshness; generate a team-brief that cross-references GitHub, Linear, Slack, and PostHog with two-way Linear sync-back; process meeting transcripts into decisions and action items with meeting-transcript; run weekly-checkin and comprehensive-analysis for cross-domain pattern analysis; launch multi-agent parallel strategic research with auto-research; and execute a full PM workflow via create-user-story, generate-prd, generate-release-notes, and publish-to-confluence. Output lands in a structured vault (00-inbox through 06-templates); People CRM profiles auto-escalate from Tier 3 (1 mention) to Tier 1 (8+ mentions or direct meeting), each observation carrying a source citation and confidence level. The memory-hygiene skill re-verifies stored facts against the live environment and stamps last_verified plus confidence, while retro audits evidence quality of completed runs.
- A product manager on a Linear + GitHub team who wants an end-to-end lifecycle tool: PRD drafting with approval gates, duplicate-checked user stories, release notes, and Confluence publishing
- An engineering lead who wants a daily team brief cross-referencing GitHub, Linear, Slack, and PostHog, with findings synced back to Linear
- An Obsidian-based knowledge worker managing daily braindumps, weekly pattern reviews, and monthly knowledge consolidation in plain Markdown
- A researcher who needs strategic questions decomposed into parallel multi-agent research threads with traceable sources
- A writer or designer who wants anti-slop copy sharpening (no-ai-slop), chart discipline guidance (data-forms), and paired UI taste skills (taste-skill for marketing surfaces, product-ui-taste for daily-use UI)
- A privacy-conscious user who wants all notes, people profiles, and intelligence stored locally and synced via iCloud rather than a cloud service
What are this agent's strengths and limitations?
- Distinctive closed-loop verification: verifiers receive only file paths — never the worker's summary — so they independently observe artifacts, and a cross-model gate catches error classes a self-review would miss
- Everything is local Markdown with no database or external servers; Git plus iCloud provide version history and multi-device sync, keeping migration and audit costs near zero
- Covers a full PM lifecycle (PRD, user stories, release notes, knowledge-base publishing) and degrades gracefully — you can start with GitHub only and add integrations over time
- Framework files are separated from your content, and updates offer per-file keep/backup/overwrite choices so upgrades never silently destroy customizations
- Team intelligence skills depend on the GitHub CLI plus MCP integrations for Linear, Slack, and PostHog — realizing full capability means configuring multiple external accounts
- The model routing is fixed (Sonnet for data I/O, Opus for reasoning), and you pay your AI agent's API costs; heavy use can produce significant token spend
- The surface spans 33 skills across 6 agent platforms with concepts like the V-model harness, tiered People CRM, and verification lanes — the learning curve is nontrivial
- Claims such as '120+ braindumps processed' and '95%+ source accuracy' are author-reported with no independently verifiable benchmarks
How do you install or deploy this agent?
- Clone the repo: git clone https://github.com/huytieu/COG-second-brain.git && cd COG-second-brain. 2. Open the folder in your AI agent and run onboarding: Claude Code —
code .then say 'Run onboarding' (discovers .claude/skills/); Cursor — open folder, say 'Run onboarding' (.cursor-plugin/ + .cursorrules); Kiro — say 'setup COG' (.kiro/powers/); Gemini CLI — rungemini, then/onboarding(GEMINI.md + .gemini/commands/); OpenAI Codex — runcodex, say 'Run onboarding' (AGENTS.md). Alternatively install via skills.sh: npx skills add huytieu/COG-second-brain. Team skills require the GitHub CLI (gh); Linear, Slack, and PostHog connect via MCP integrations and can be added incrementally starting from GitHub only. Optional Git sync, iCloud, and Obsidian Tasks setup is documented in SETUP.md. Total setup is roughly 2 minutes.
How do you use this agent?
Run onboarding once, then invoke skills in natural language: 'I need to braindump' to capture thoughts; 'Give me my daily brief' for news intelligence; 'Team brief' for team intelligence; 'Process this meeting' for transcripts; 'Weekly review' for pattern analysis; 'Run this through the closed loop' to put a build task through the V-model verification lifecycle; 'Harvest what we learned' to stage session learnings for approval as skill patches; 'Audit my memories' for a memory trust sweep. Update the framework with 'Update COG' or ./cog-update.sh (supports --check, --dry-run, --force); check the version with cat COG-VERSION. Validate agent-surface consistency with ./scripts/validate-agent-surface.sh before publishing framework changes.
How does this agent compare with similar options?
The README positions COG against Notion, Roam, and plain Obsidian, which only store; COG adds self-evolving classification, pattern analysis, and consolidation on top of local Markdown. Its agent architecture is directly inspired by Garry Tan's gstack (specialist sessions, model routing) and gbrain (tiered people-profile enrichment, knowledge-first lookup), and it incorporates Zettelkasten, PARA (Tiago Forte), and GTD methodology.