install

/matts-second-brain:install — wire Trilium up as Claude's brain

Walks through the full setup. Idempotent — safe to re-run.

Phase 1 — preflight

Check what's already in place. Report each as [ok] / [needs setup]:

# Trilium reachable?
curl -sf -o /dev/null "${TRILIUM_URL:-http://127.0.0.1:8787}/api/health-check" \
  && echo "[ok] Trilium reachable" \
  || echo "[needs setup] Trilium not reachable"

# trilium-mcp jar built?
test -f /home/matt/mcp/trilium-mcp/target/trilium-mcp-1.0.0.jar \
  && echo "[ok] trilium-mcp jar present" \
  || echo "[needs setup] trilium-mcp jar missing"

# Token present?
test -s /etc/claude-brain/trilium-token \
  && echo "[ok] ETAPI token configured" \
  || echo "[needs setup] ETAPI token not set"

# MCP registered in user's Claude config?
grep -q '"trilium"' ~/.claude.json 2>/dev/null \
  && echo "[ok] trilium MCP registered" \
  || echo "[needs setup] trilium MCP not registered"

Phase 2 — install missing pieces

For each [needs setup], walk the user through it interactively. Don't run destructive things without confirmation.

If Trilium isn't reachable

Offer to bring up a docker container:

• compose snippet under docs/trilium-docker-compose.yml
• run docker compose up -d
• wait for healthcheck

If trilium-mcp jar missing

• Clone https://github.com/wrxck/trilium-mcp if not present in /home/matt/mcp/trilium-mcp
• Run mvn -q package inside it
• Verify the jar landed

If token not set

• Tell the user: "Open Trilium → Options → ETAPI → Create new token"
• Prompt for the value with hidden input (Bash read -rs)
• Write to /etc/claude-brain/trilium-token with mode 0640, owner root:claude
• NEVER echo the token back, NEVER include it in any audit/log line

If MCP not registered

• Append the trilium MCP entry to ~/.claude.json (or merge with existing mcpServers)
• Use /usr/local/bin/trilium-mcp as the command — wrapper handles env injection

Phase 3 — seed the brain taxonomy

Use trilium MCP tools to create the root note + structure if not already there:

Claude Memory/
  ├── 00 — How to use this brain     # discipline rules (see skills/_discipline)
  ├── Standards/
  ├── Decisions/
  ├── Lessons Learned/
  ├── Apps/
  ├── Reviews/
  └── Drafts/

Tag every note with #claude-brain. If the root already exists, skip — never overwrite the user's existing notes.

Phase 4 — multi-user (opt-in)

If --multi-user flag passed:

• Create claude group if missing
• Add current user + root + any user passed via --users a,b,c
• Move/copy token to /etc/claude-brain/ with group ownership
• Drop /etc/profile.d/claude-brain.sh so future login shells auto-register the MCP
• Symlink /usr/local/bin/trilium-mcp if not present

Phase 5 — transcript scan (opt-in)

If --scan-transcripts flag passed:

• Walk ~/.claude/projects/*/ for .jsonl session transcripts
• Run brain_scan_transcripts MCP tool to infer:
  • Repeated user corrections → candidate Standards/
  • Resolved bugs with rationale → candidate Lessons Learned/
  • Per-app gotchas → candidate Apps/<name>/
• Always present inferred notes for human review before writing — never bulk-create. The user accepts/rejects each.

Phase 6 — verify

End by:

1. Calling brain_recall("git workflow") — should return citations from the seeded standards
2. Telling the user to quit + relaunch Claude (so the new MCP registration takes effect)
3. Suggesting first manual note to write so they get a feel for brain_remember