Hermes Agent

Hermes Agent is the personal AI agent by Nous Research — “the agent that grows with you”. Open source (MIT, hermes-agent (OSS)), runs entirely on your own machine, and is model-agnostic: any OpenAI-compatible endpoint works, local or cloud. It exposes itself through three faces — terminal CLI, messaging platforms (Telegram, WhatsApp, Slack, Discord, Matrix, Mattermost), and a desktop app — all backed by the same agent core with tools, skills, memory, cron, and subagents.

The design bet: instead of a cloud assistant you rent, a long-lived local process that accumulates skills and memory, checked into version control, wired into your infrastructure.

This deployment (doombox)

• Install: v0.21.3, git method, ~/.hermes/hermes-agent on main at upstream db1f3f45 — plus one local patch branch local/doombox-patches carrying a single commit: configurable STT HTTP timeout (hard-coded 30 s was too short for CPU Parakeet; the patch reads stt.openai.timeout, default 300 s; PR-ready).
• Platform: Telegram — this note is being dictated into a Telegram thread right now. 257 sessions on disk.
• Model: provider zai (GLM) for the main profile; a second profile iris routes to llama-swap (OSS) on the homelab GPU for local inference.
• Voice loop, all local: STT via voiceapi (Parakeet, :5092, 300 s timeout above); TTS via omnivoice (:6655, voice “filip”) through the OpenAI-compatible provider config.
• Extensions: 41 skills (nested categories from devops to gaming), agentmemory plugin, qmd MCP server for hybrid search over the vault.
• Update cadence: 1,099 commits behind upstream — deliberately lagging; updates are manual hermes update runs.
• Config discipline: edits only via hermes config set/unset; direct config.yaml writes are blocked by policy.

The interest here is less the agent itself than the pattern: a durable agent process colocated with the services it uses (STT/TTS/inference all on LAN loopback), with the Local LM roadmap as the wider plan it lives in.

Related captures

• @witcheer’s context-file audit prompt — a copy-paste prompt that makes a Hermes instance audit its own SOUL.md/memory/AGENTS.md line-by-line against the live machine (captured privately in clippings/x-2102713441316168048-thread.md). Not run on this instance — declined.
• @witcheer’s MY-HERMES.md manual prompt — the sibling idea: the agent writes a one-page rebuild manual of itself (model, provider, platforms, cron, skills, config/memory/session paths, backup/restore), verified live with read-only commands, secrets shown only as “set”/“not set”, draft-shown-before-write (clippings/x-2103764777621147741-thread.md). Same author, same trust pattern — self-documentation instead of audit.
• The ecosystem around it (same org): hermes-agent-self-evolution (DSPy+GEPA), Bot Mode, paperclip-adapter — captured in the org listing, not individually.

Links

• hermes-agent (OSS)
• llama-swap (OSS)
• Speech to text inference
• Local LM roadmap
• QMD

Profile distributions (the packaging story)

hermes-star-trek-profiles (OSS) (teknium1 — Nous Research’s own Teknium; 113★; via @HermesWatcher, 670❤/43k views) — 68 installable Star Trek personas (TOS/TNG/DS9/VOY): Spock = research + root-cause; Picard = strategy + executive decisions; Data = technical/data analysis; Worf = security + risk; Quark = pricing + negotiation. The mechanism is the interesting part: each profile = a substantial SOUL.md (identity, voice, worldview, operating method, strengths, blind spots, pressure behavior, disagreement style, safeguards, task affinities) + character terminal skin + provider-neutral config + distribution.yaml; ships no credentials, memories, sessions, model choice, cron, or MCP servers (“your provider setup remains yours”); manage.py install one profile or a whole series. The watcher’s takeaway is the right one: “Build the role once. Package it. Share it. Let anyone install it.” — personas as portable distributions, the same move as skills/packages in other harnesses. Also proof the profile system (this session’s iris = local llama-swap backend) is a distribution format, not just a local config fork.

The community cheat sheet (2026-10-05)

HermesWatcher (190❤ / 10k views) — the Hermes command cheat sheet, new edition: fresh layout, per-command badges showing where each works (Desktop / CLI / messaging / shell), and — the part that makes it durable — “command names, syntax, and availability checked against Hermes source.” Covers /goal, /loop, /heartbeat, /bg, /btw, /plan, /review, /refine, /moa, /busy, plus hermes pause/hermes resume in Terminal Essentials, expanded CLI queue guidance, and branching/saved-sessions clarification. Organized by intent: sessions & context · active work · goals/loops/plans · models & behavior · skills/memory/tools · automation & coordination · inspection/recovery/troubleshooting. Second HermesWatcher artifact in the vault (after the Star Trek profiles) — this account is becoming the de-facto community documentation lane.

Hermes on Kubernetes: the hermeum operator (2026-10-08)

learnk8s → hermeum/hermes-agent-operator (42★ Go; captured): “Hermes Agent is a powerful tool — but designed for personal use. Running it across a team is difficult: configurations drift, skills go out of sync, and there is no shared source of truth for what each agent does or what it has access to.” The operator declares the full agent state — config, workspace files, skills, crons, bundles, plugins, packages, resources, env — as a single HermesAgent CR: the operator reconciles the running agent against it. The recursive bit: a Hermes skill (hermes skills install hermeum/hermes-agent-operator/...) lets a running Hermes agent scaffold and apply manifests for new agents — agents deploying agents via CRs. Parent platform: Hermeum (dashboard, templates, shared credentials) built on the operator. For this note’s ecosystem lane: the K8s-operators-for-harnesses pattern lands on Hermes itself — the same declarative-cure-for-drift argument as Rewriting Ansible roles from K8s operators (inverted: operator wrapping the agent instead of the role), and the Meta harness enterprise pole’s missing infrastructure layer. Homelab-relevant: declarative Hermes on k3s is exactly the doombox-shaped pattern the Homelabs candidates ledger tracks.

The witcheer video canon: teams-of-agents section → playlist (2026-10-08)

witcheer (122❤; thread + the original Sep-28 map tweet both captured): the answer to “how do I run a team of Hermes agents” — the fifth section of his video map, now a YouTube playlist: multi-agent setups with Kanban, Bot Mode and cron jobs. The full map (from the quoted tweet): (1) start here · (2) running a business on it · (3) real setups and use cases · (4) skills · (5) teams of agents · (6) by job: marketing, sales, support, research. Third witcheer artifact in the vault (after the self-audit prompt and MY-HERMES.md manual) — the pattern holds: operationally-dense Hermes content, checklist-shaped. The teams-of-agents question is the same one the Meta harness note tracks from the product side; witcheer’s playlist is the practice side — Kanban lanes + Bot Mode + cron as the poor-man’s orchestration before any meta-harness is needed. Playlist: youtube.com/playlist?list=PLEf1uH3e1AUo.