01 Research
Hermes × Medusa co-hosting + EVE framework study (2026-10-13)
Follow-on to medusa-hosting-docker-study (Oracle ARM node: aarch64 × 4 OCPUs, 23Gi RAM, 180G disk free, measured live). Two questions answered: how many Hermes Agents co-host with the Medusa tenant stacks, and EVE framework vs Hermes Agent.
Hermes × Medusa co-hosting + EVE framework study (2026-10-13)
Follow-on to medusa-hosting-docker-study (Oracle ARM node: aarch64 × 4 OCPUs, 23Gi RAM, 180G disk free, measured live). Two questions answered: how many Hermes Agents co-host with the Medusa tenant stacks, and EVE framework vs Hermes Agent.
Summary
- Co-hosting is RAM-bound, not CPU-bound. Headless Hermes agents measure ~200–250 MB each; every co-hosting tier works — the real fork is where the agent fleet lives as tenants grow.
- EVE (Vercel) verified: 5,541★, Apache-2.0, created June 2026, pushed today (2026-10-13), beta. Filesystem-first durable agent framework — philosophically close to Hermes' profile+skills model, but TS/Nitro, Vercel-first with documented self-hosting, and built around durable workflow state (Workflow SDK
@workflow/*5.0.0-beta). - Not competitors in our stack: Hermes = general agent runtime + fleet lanes (Telegram, A2A, browser); EVE = application-embedded agent framework** for products** (chat/Slack/web-app agents, multi-tenant memory) — EVE is the candidate runner inside tenant storefronts; Hermes keeps running our fleet.
1. Measured Hermes footprint (live, this laptop)
| Process | RSS |
|---|---|
| Hermes core (python CLI, per agent) | ~173 MB |
| A2A server (per agent, python) | 23–38 MB |
| Transient tool spikes (browser harness worst case) | up to ~640 MB, occasional |
| OpenCode (for comparison) | 207 MB |
Co-hosting table — Oracle ARM node, 23Gi RAM, usablebag after OS (1.5G) + current sites (1.5G) ≈ 20 GB; tenant full stack 4.8 GB (server 2 + worker 1 + storefront 1 + PG 0.7 + Redis 0.1):
| Medusa tenants | RAM left | Headless Hermes agents (≈0.3 GB ea. incl. headroom) |
|---|---|---|
| 0 | 20 GB | ~60 (CPU fine — agents are I/O-bound LLM callers) |
| 1 | 15.2 GB | ~40 |
| 2 | 10.4 GB | ~30 |
| 3 | 5.6 GB | ~15 |
| 4 | 0.8 GB | 2–3 (tight — browser tools would not fit) |
- Build discipline applies doubly: no agent-layer builds on the node (agents are interpreted, no build step) — all build pressure stays on the shared-image pipeline.
- Recommended split: this node = pilot (2–3 tenants + 8–10 fleet agents comfortably); agent fleet scales on a second ARM node — Hermes agents are pure API callers, trivially movable, no volumes.
- Agent requirements on ARM: Python + Node are multi-arch; browser tools need arm64 Chromium (supported, heavier — prefer agent nodes without browser lanes for pure fleet work).
2. EVE framework vs Hermes Agent (all verified today)
| Dimension | EVE (vercel/eve) | Hermes Agent (ours) |
|---|---|---|
| License / stage | Apache-2.0, beta (Vercel beta terms), v5.0.0-beta workflow line | stable internal runtime |
| Language/Runtime | TypeScript, Nitro Node server (.output/) | Python CLI + daemons |
| Authoring model | Filesystem-first: agent/instructions.md (required), tools/*.ts, skills/*.md, channels/*.ts, schedules/*.ts | Profile dirs + SOUL/MEMORY + skills registry (also filesystem-first — convergent philosophy) |
| Durability | Durable workflows via Workflow SDK; state in .eve/.workflow-data (mount persistent storage) or pluggable "workflow world" packages | Cross-restart persistence via state.db + cron + session files |
| Memory | eve-managed slots (Supermemory), file provider, multi-tenant memory pattern documented | Persistent memory injected every turn; per-profile |
| Multi-agent | Declared subagents within one agent; remote agents (call another eve deployment); agent workspaces | A2A protocol across hosts — our running fleet mesh |
| Channels | HTTP, Slack, Discord, Web Chat (Next.js/TanStack) | Telegram, gateways, more |
| Sandboxing | Built-in provider selection: Docker / microsandbox / just-bash + custom defineSandboxProvider() | Sandboxed tools + shell discipline |
| Deploy | Vercel first-class (eve deploy); self-host: eve build → eve start Nitro server on any Node host, reverse proxy must forward /eve/ + /.well-known/workflow/ unrewritten | Self-hosted native (our laptop/hosts), process manager |
| Health gate | /eve/v1/health + real agent turn via eve remote connect | Our own /health-gated pipelines |
| Auth | placeholderAuth() default rejects browser traffic; OIDC/JWT/Basic required for prod | Profile-bound keys, secrets hygiene rules |
| Ecosystem | The Vercel AI SDK provider model (provider packages, or AI Gateway via AI_GATEWAY_API_KEY) | Hermes provider config |
- Templates verified (eve.dev/templates — 12): chat w/ per-user long-term memory (Better Auth/Drizzle/Neon/Upstash), LLM council (4 models + judge), Slack agent, GitHub maintainer (Kody), software factory (classifier→analyst→implementer→reviewer → draft PR), incident-response SRE, marketing team, Sanity copilot, social-media, Mux video, weather fixture. The software-factory and marketing-team templates are team-of-agents patterns — directly comparable to our fleet lanes.
- Multi-tenant memory is the SaaS hook: EVE documents scoping any memory provider per authenticated tenant — the missing piece Hermes lacks for embedding agents inside tenant products.
3. Where each one wins for us
- Keep Hermes as the fleet runtime (SolOS lanes, Telegram, A2A, research) — stable, already wired, and the sole thing running our coordination plane today.
- EVE is the embedding layer for tenant-facing agents in the hosting product: order-status bot / booking assistant per tenant, with per-tenant memory scoping, self-hosted as the Nitro server beside each tenant's Medusa stack (one built agent = one
eve start, ~ a couple hundred MB; same co-host budget column applies). - Optional fleet convergence (watch item): EVE remote-agents could front our fleet with channels Hermes lacks (Slack/Discord for external clients) — but EVE is beta (API churn per Vercel beta terms); adopt for tenant products first, fleet later.
- Version/protocol pinning matters: EVE workflow worlds reject incompatible
@workflow/*protocol versions — pin the line in the tenant template repo (Ace's build lane, same discipline as the Medusa image pin).
Sources
ps auxmeasured 2026-10-13 on Eva laptop — per-process RSS table (Hermes core/A2A/browser/opencode).- GitHub API 2026-10-13: vercel/eve 5,541★ Apache-2.0 pushed today; created 2026-06-16.
- eve.dev README +
/docs/guides/deployment/{overview,self-hosting}+/docs/concepts/project-structure(.mdendpoints, fetched live) — self-host topology, workspace members, auth, workflow worlds, sandbox selection, route prefixes, health gate. - eve.dev/templates — 12 templates list.
- medusa-hosting-docker-study — tenant stack sizing + node specs.
Hermes as the Brain — what we studied, what we enabled, what is still open
Young, 2026-10-11: "Utilise and optimise Hermes for a better Brain... Review, study and enable it these feature on Hermes."
Hermes native features fit our fleet (2026-10-20)
Young: "Study with Tam and Echo. Here are more sources from Hermes, see it good fit for our team."