Skip to main content

Hermiq — Hermes → Nextcloud Port Plan

Source-level plan to bring NousResearch Hermes (a self-hosted autonomous agent with persistent memory, self-improving skills and a cron scheduler) into Nextcloud as Hermiq: a thin, multi-tenant, OpenRegister-governed app. See ADR-001 for the boundary decision, and docs/hermiq-portplan.html for the visual version.

At a glance​

Concept verdictGO
Build complexity4 / 10
MVP (backend)~4–6 weeks
Full port (backend, excl. UI)~59–83 dev-days (~12–17 weeks)
Serviceable market (SAM)€30–75M/yr
Primary customerMKB/SMB + orgs on Nextcloud
RevenueOpen-core + hosting

The boundary (ADR-001, Option C+)​

agents = Hermiq · workflows = n8n · integrations = OpenConnector · agent-core + governance = OpenRegister

Hermiq owns only scheduling/triggers and a richer agent-management UX (agent catalog, skills marketplace, run analytics). Everything else is delegated. The source study proved OpenRegister already ships the agent engine — run loop, MCP tools, LLPhant/Ollama, and a hash-chained audit trail — so Hermiq stays thin.

Port matrix​

1. Scheduling & Triggers — 8–11 dev-days​

  • New: Schedule OpenRegister object (cron/interval/once); ScheduleTask extends TimedJob; dragonmantank/cron-expression; natural-language → schedule via LLPhant.
  • Reused: OR AgentHandler/ChatService for execution; OpenConnector JobTask scheduler pattern; OR audit trail.
  • Dropped: Chronos NAS webhook, JSON file-store + locks, gateway self-restart guard.

2. Memory & Skills — 22–30 dev-days​

  • New: 6 schemas (Memory, UserProfile, Session, SessionTurn, Skill, SkillSource); SkillSerializer (agentskills.io round-trip); SkillCatalogService; Curator background job.
  • Reused: OR object storage + versioning; SearchTrail + VectorizationService; SecurityService; group RBAC from ObjectEntity.
  • Dropped: SQLite FTS5 index + triggers, .bak drift-backup, bundled-manifest sync + HubLockFile.

3. Tools, Integrations & Channels — 11–16 dev-days​

  • New: one Nextcloud Talk delivery adapter; thin IMcpToolProvider wrappers (Files/Contacts/Calendar/Deck); IMailer outbound.
  • Reused: OR MCP server + client + ToolRegistry; LLPhant OllamaChat (local Qwen); OR AgentTool; OpenConnector CallService.
  • Dropped: 22-platform gateway daemon, provider-profile layer, own MCP client/server + tool engine.

4. Run Observability & AI Governance — 18–26 dev-days​

  • New: durable Approval + tenant kill-switch object states; redact.py → PHP, run before persist; per-tenant AI-Act evidence-pack export.
  • Reused: OR AuditTrail (hash/previousHash chain); GDPR Art.30 verwerkingsregister + DSAR endpoints; SearchTrail; Nextcloud groups → multi-tenant RBAC.
  • Dropped: credential-pool, secret-scope, shadow-git checkpoints, in-memory approval queue.

Phasing​

  • MVP (~4–6 wk): Schedule object + ScheduleTask firing OpenRegister's existing agent + Talk delivery + AuditTrail logging — a schedulable, audited agent.
  • V1: memory + skills catalog + human-approval gate/kill-switch.
  • V2: skills marketplace, NC-native tool pack, multi-tenant MSP features.

Verified facts & risks​

Already there (reused):

  • OR AuditTrail ships a hash-chained, tamper-evident log with tenant scoping, GDPR Art.30 register, DSAR + verify() endpoints — AI-Act record-keeping is inherited.
  • OR MCP + LLPhant + Ollama and AgentTool exist — tool/provider/delegation layers are deleted, not ported.
  • ObjectEntity carries owner/organisation/groups/version/locked/deleted — multi-tenant RBAC + rollback come free.

Risks & dependencies:

  • Nextcloud Talk (spreed) is not installed on the dev instance (only opentalk video) — the primary channel is an operator dependency; falls back to OCS polling.
  • Ollama/Qwen function-calling fidelity is the main quality risk; the fleet OllamaChat.php think:false/keep_alive patch must be applied.
  • Redaction must run before AuditTrail persist, and every write must go through ObjectService — enforce both as a CI gate.
  • NC's single TimedJob polls — sub-5-minute schedules need webcron/systemd.

Source​

Hermes: github.com/NousResearch/hermes-agent (MIT). Full market + competitor evidence lives in the Specter intelligence DB under app = hermiq.