title: Claude Code workflow suggestions — round 1 date: 2026-05-24 host: ONE (10.10.1.120) scope: homelab ops/infra, vault knowledge mgmt, agent infra/memory risk: wide net (stable + experimental)
Companion analysis of your current Claude Code setup with curated additions
targeting your three stated focus areas. Items marked [exp] are
experimental/newer; [stable] are battle-tested; no marker means active but
not yet 1.0.
13 plugin marketplaces installed — already very rich. You have:
| Category | Already covered by |
|---|---|
| Memory/observability | claude-mem (timeline, knowledge-agent, weekly-digests, oh-my-issues, babysit, design-is, pathfinder, version-bump, smart-explore, wowerpoint) |
| Markdown tooling | han/markdown (just fixed — marksman LSP), claude-mem:make-plan/do |
| Karpathy guidelines | andrej-karpathy-skills |
| Vault automation | 50+ custom /vault-* commands (your own) |
| Dev infra | claude-dev-infrastructure (qa, plugin-creator, ai-truthfulness, etc.) |
| Sub-agents | voltagent-subagents (35 agents loaded) |
| LSPs | claude-code-lsps (4 LSP servers, incl. marksman now) |
25 MCP servers configured — heavy investment in: cloudns, cognee, context7, docker-mcp, fetch, filesystem, firecrawl, lean-ctx, nmap, obsidian, pfsense, plesk, postgres, powershell-ssh, proxmox, puppeteer, qdrant-memory, sequential-thinking, shell, silverbullet, sqlite, ssh, web-reader, web-search-prime, zread.
Pain points surfaced from mem-search (last 30 days):
- Cognee Gemini max_tokens truncations on cognify runs (#3589, #3590)
- GRAPH_COMPLETION search timeouts via MCP 3s client limit (#3550)
- SMB file enumeration timeouts after 120s (#3334)
- Repeated MCP transport tracebacks corrupting poller loops (#3512)
- Plugin LSP install scripts assuming brew/cargo only (today's marksman issue)
- ClouDNS MCP didn't exist — you wrote your own (#2882, #2886)
- Pre-cache claude-mem corpora before query, else null (already noted in CLAUDE.md)
[stable] — official; richer than the gh CLI wrapping in many tools. Push/PR/issue automation across all your repos.[exp] — if you run HA. Voice into device/script automation.[exp] — peer mgmt for inter-host tunnels.[exp] — if you spin up a small k3s on Proxmox.mcp-server-pfsense for write ops, or fork yours to add the missing mutations.[exp] — playbook execution from Claude.vault-fact-check.[exp] — log queries for Grafana Loki.[exp] — monitor-status into Claude.[stable] — you have an obsidian MCP; this one is REST-via-local-API plugin, often richer than file-only access (search-by-property, dataview-aware).[exp] — alternative graph backend; could replace FalkorDB you already run, and Memgraph has stronger Cypher support than Cognee's current backend.[exp] — embeddings/RAG without the cognee pipeline weight; good for vault-only Q&A.find for vault lookups.[exp] — if you've ever considered Anytype as a structured-vault complement.[exp] — pull research papers into vault Knowledge/Research/.[stable] — alternative memory layer with multi-user/agent scoping; complements openmemory-1 you already run.[exp] — MemGPT successor; long-term episodic memory with summarization tiers.[stable] — temporal knowledge graph for conversation memory; tighter than cognee for chat history specifically.[stable] — you already run graphiti/ docker stack but no MCP front yet; the official MCP would let Claude write into it.[exp] — coordinate sub-workflows that span multiple Claude calls.[exp] — community prompt collections as installable.[exp] — multi-instance Claude Code orchestration.[exp] — ephemeral docker sandboxes for risky commands; pairs well with your "Before Destructive Actions" rule.[exp] — agentic workflow framework on top of Claude Code; SPARC patterns, swarm modes.You already have an extensive /vault-* family. Gaps worth filling:
/doctor output with auto-remediation hints (you essentially did this manually today for marksman).docker-validate skill: check image freshness, recreate-pending, env-drift between compose files. You have 25+ stacks — drift is inevitable.sync-wsl-firewall scripts; wrap them as a skill so Claude can self-serve new port exposures.[stable] — Retrieval-as-a-Service; fastapi + postgres + graph layer; arguably more reliable than current cognee setup for production-style RAG. Has an MCP server.[exp] — multimodal RAG (PDFs, images, video); strong on document understanding.[stable] — RAG UI + backend; could replace cognee UI.[stable] — enterprise-grade search over vault + Slack + GH + Confluence; runs as docker stack.[exp] — Obsidian-compatible vault with built-in local LLM + auto-linking; good companion to Obsidian on the editing side.[exp] — NotebookLM open-source clone; might replace wowerpoint skill for richer deck output.[stable] — lightweight monitoring for docker containers + hosts; one binary per server + central dashboard. Better fit than full Prometheus stack for homelab.[stable] — real-time docker log viewer; complements your existing nginx-proxy-manager admin UI.[stable] — htop on steroids; web UI mode + REST API → pair with a glances-mcp wrapper.[stable] — you already have /opt/komodo; if not yet active, it's an excellent docker orchestration UI across multiple hosts.[stable] — you have portainer dir; agents on zeus/ajax-proxmox give single-pane container management.[exp] — NetBox MCP if you adopt it for IPAM; would replace your hand-maintained Network/10.10.1.0-24-Hosts.md.[stable] — public ingress without opening pfSense ports; complement to *.qoop.au NPM.[stable] — self-hosted Tailscale control plane; if you want full sovereignty over the WireGuard mesh.[stable] — modern alternative to your MkDocs publisher; native Obsidian wikilinks, faster builds, better search. Worth evaluating against the /root/vault-publish pipeline.[exp] — direct from-Obsidian static export; may sidestep your pygments/SilverBullet crash hunt entirely.[exp] — PHP-based Obsidian web reader; lightest possible publish path.[stable] — if you want a docs-site-quality publish for selected vault subtrees.[stable] — you have /opt/litellm; if not active, it's the canonical multi-provider router (Gemini/Anthropic/Ollama all behind one OpenAI-compatible endpoint).[exp] — persistent agent server; could replace mcp-memory + openmemory consolidation.[stable] — git-native AI pair programmer; lighter than Claude Code for quick batch refactors.[exp] — long-running task agent with branch-style sandbox; complement to claude-mem:make-plan.| Pain point (obs ID) | Suggested addition |
|---|---|
Cognee Gemini max_tokens truncations (#3589) |
Add litellm proxy in front of cognee; set max_tokens ceiling + automatic chunking + fallback to local Ollama on truncation |
GRAPH_COMPLETION MCP 3s timeout (#3550) |
Wrap cognee MCP with a retry+REST-fallback shim (you already do this manually for claude-mem) |
| SMB file enum 120s timeout (#3334) | Cache vault tree to a local SQLite mirror; install Everything mcp or use fd -L w/ exclude rules |
| MCP transport tracebacks corrupting poller (#3512) | mcp-supervisor pattern: PM2/supervisord watching every MCP stdio, restart on health failure |
| ClouDNS had no MCP, you built one (#2886) | Publish your cloudns-mcp to the marketplace; check pfsense-mcp write-ops branch too |
| Plugin install assumes brew/cargo (today) | plugin-doctor skill that auto-detects platform + installs binaries; PR back upstream to han/markdown |
| Repeated "is process X already running?" checks | Add Beszel or Dozzle for at-a-glance container/process state; reduce bash probing |
~/.local/bin → /usr/local/bin) is a recurring pattern. Add a SessionStart hook that ensures /usr/local/bin has symlinks for all ~/.local/bin/* binaries plugins might need./mcp-health slash command — 30-line bash that hits every MCP's known health endpoint; would have caught marksman LSP within 1 cycle.vault-cognee-sync skill — incremental cognee re-ingest only for vault files changed since last cognify run; avoids the 40+ minute full rebuilds./doctor autorun on SessionStart — surface plugin issues immediately, not next time you remember.mem-search corpus prewarm cron — daily background prime so you never hit the null-on-cold-cache trap.~/.claude/local-bin-bootstrap.sh.Resist the urge to install: - Another vector DB — you have qdrant (×2), falkordb, cognee's internal. Consolidate, don't add. - Yet another memory layer — you have claude-mem, openmemory, mcp-memory, qdrant-memory, cognee. The architecture sprawl is itself a maintenance load. - More UI dashboards — silverbullet, hermes, vault-site, openwebui, cognee-ui, NPM admin, portainer (if active). Each is a small upkeep tax.
When evaluating any of the above sections, the question to ask is: does this replace something already running, or just add to the inventory? Replacement proposals are higher-value than additions at this stage.
/opt/komodo) actively used? If yes, several entries above are redundant. If no, it's a strong candidate to centralize the 25-stack management./opt/litellm) running? Several pain-point fixes depend on it being the front for Gemini/Ollama.Next round (suggestions-2.md) can drill into whichever of these you want to
explore. Just say which sections to expand or which open question to resolve
first.