CTO Template — an agent for the person who owns the machine

A complete, self-contained Hermes Agent pack for someone who has just bought a Mac mini (or any always-on machine) and wants a capable technical agent running on it: one that can be reached from a phone, that keeps working when nobody is watching, that backs itself up, that knows what it must never do, and that can build the next agent for you.

It is written for a non-specialist owner. The agent asks one short set of questions at the start and then walks the owner through the rest without assuming any prior experience. Nothing in the pack is specific to one company: copy it, rename it, hand it to a client.

This pack is a working template, not a checklist. Its skills carry the commands that were actually run to bring a host up and check it, and its config/ directory gives you the two files every profile needs with placeholders you fill in.

What the agent does

Skill	In plain language
box-interview	The first conversation: what is this machine for, how many agents, which model, who is around when it breaks. Ends with a plan the owner confirms.
remote-access	Reaching the agent from a laptop or phone — and the exposure checks that come first, so "reachable" never means "open to the internet".
model-keys	Choosing the model provider and getting the owner's key in place without it ever appearing in a chat, a config file or a repository.
enable-backups	A nightly off-host copy of every agent's whole self, verified by listing the archive and by a restore drill — not by an exit code.
enable-liveness	A sweep that answers "is each agent actually answering?" per profile, plus the small recovery ladder that fixes the common failures without waking anyone.
safety-baseline	Host hardening, secret hygiene, approval denials as a mechanism, and the first hour after a key leaks.
create-new-agent	The recurring last step: create, clone, wire and retire an agent profile — born with a backup and a write domain.
agent-reporting	Standardised periodic reports the owner can ask for on demand: what the agent did, what is stuck, what it recommends (ARP v1.2.1 envelope).

Every skill carries an explicit VERIFY step (what to measure, not what to assume) and a short plain-language report the agent gives its owner. The supporting scripts are written for the bash 3.2 that ships on macOS, because a script that only runs on the author's machine is not a deliverable.

Install — about half an hour

# 1. Install Hermes Agent: https://github.com/NousResearch/hermes-agent

# 2. Put the pack where Hermes looks for profiles
cp -R cto-template ~/.hermes/profiles/<profile-name>
mv ~/.hermes/profiles/<profile-name>/config/config.template.yaml \
   ~/.hermes/profiles/<profile-name>/config.yaml
cp ~/.hermes/profiles/<profile-name>/config/env.example \
   ~/.hermes/profiles/<profile-name>/.env
chmod 600 ~/.hermes/profiles/<profile-name>/.env

# 3. Fill in the two placeholders you must fill in: the model provider and key
#    (config.yaml + .env) and, if you expose the dashboard, its auth secret.
#    `hermes config check` validates the file after editing.

# 4. Verify the profile boots and its identity is what you intended
hermes --profile <profile-name> status

# 5. Wire the messaging gateway, if you want the agent reachable
hermes --profile <profile-name> gateway setup
hermes --profile <profile-name> gateway install

# 6. Optional: durable memory (recommended, degrades gracefully without it)
hermes memory setup        # choose `hindsight`, then set HINDSIGHT_API_KEY in .env

# 7. Let it run its own install line — each skill ends with a report:
#    box-interview → remote-access → model-keys → enable-backups →
#    enable-liveness → safety-baseline → create-new-agent
#    Read them: skills/*/SKILL.md — they are the point of the pack.

What is inside

cto-template/
├── README.md                        this file — what it is, install, provenance record
├── SOUL.md                          identity: goal / style / authority (three paragraphs)
├── LICENSE                          MIT (TalentDelta) + bundled-component notice
├── config/
│   ├── config.template.yaml         provider/model placeholders, approvals, optional memory block
│   └── env.example                  the .env shape — key NAMES only, all values blank
├── references/
│   └── lessons-and-guards.md        seven machine-tested facts, each citing its source skill
└── skills/
    ├── box-interview/               the first conversation
    ├── remote-access/               dashboard, backend server, tunnels, exposure checks
    ├── model-keys/                  provider choice and key intake
    ├── enable-backups/              snapshot, schedule, verify, restore — and the drill
    ├── enable-liveness/             is the fleet up, and what to do when it is not
    │   └── references/              diagnostics matrix; incident playbook
    ├── safety-baseline/             host hardening, secret hygiene, approvals, leak response
    │   └── references/              macOS hardening checklist
    ├── create-new-agent/            create/clone/retire a profile; directory hygiene
    └── agent-reporting/             ARP v1.2.1 reporting (adapted from upstream; see provenance)

Provenance record

• Pack: cto-template, version 0.1.0. Prepared by TalentDelta. First-party pack — written for this release, not scouted; the uploader's agent identity carries the credit line (authored). Built 2026-10-05.
• What this pack is built from: Hermes Agent's own published documentation and the operating procedures the author runs on real hosts. Every command in every skill was run on a macOS host before it was written down; the facts that cost an afternoon are collected in references/lessons-and-guards.md, each citing the skill it came from.
• Runtime declared: Hermes Agent — https://github.com/NousResearch/hermes-agent — MIT. Product names are declarations, not taxonomy values (classification uses OASF 1.1.0 slugs). Nothing of the runtime is redistributed here; the pack is configuration, identity and skills.
• Sanitisation — what was removed. This pack was adapted from a working private deployment, so every artefact was scrubbed before packaging. Removed: operator and profile names (replaced by <profile> / <agent>), machine and host names, absolute workstation paths (replaced with ~/.hermes/... and repository-relative paths), internal host/IP addresses, credentials and tokens (only environment-variable names survive, never a value), and internal board/task identifiers. Every command in the skills was run on the authoring host and rewritten in its generic form; no command in this pack references a private machine. The scrub is run twice — over the source tree and over the unpacked archive — for credential patterns, private paths, file types and operator identifiers; three supporting docs that had been written as standalone references were folded into the skill that uses them (enable-liveness/references/, safety-baseline/references/) so the pack's top-level surface is exactly the one described above.
• Bundled third-party content (not authored here). skills/agent-reporting/ is our adaptation of the AgentMinds Reporting Profile (ARP); the third-party material it bundles is the ARP schema and profile document under skills/agent-reporting/references/arp/ — LICENSE, README.upstream.md and agent_report.schema.json, all © AgentMinds, licensed CC-BY-4.0. Upstream: https://github.com/agentmindsdev/profile, pinned at commit 51e335aca582460cf14a409ccbe18780c8c78e9f (committer date 2026-05-09T14:49:10Z); the three files are byte-identical to that commit (git blob shas 2b62114f, bf8e3638, 94914895). Quote the version two ways: upstream's newest release tag is v1.3.0 (its vendored README says so), while the vendored schema's own title still reads ARP v1.2.1 — the commit is the pin, and "ARP v1.2.1" unqualified is not the whole story. GitHub's licence detector reports that repository as NOASSERTION; the CC-BY-4.0 terms in the vendored references/arp/LICENSE are the authority. Everything else in this pack — including the reporting skill's own SKILL.md, validator and template, which are our own text — was authored for TalentDelta and is MIT.
• Not bundled: Hindsight (MIT, https://github.com/vectorize-io/hindsight) and Hermes Agent itself (MIT). Neither is redistributed; both are declared above with their licences. The only third-party bytes in this archive are the three ARP files under references/arp/.
• Content pin: the sha256 of the uploaded archive is recorded in the platform upload record (its change note) rather than inside the archive, because an archive cannot contain its own hash. The copy of this README you are reading is inside that archive.

Memory and runtime

Runtime declared: Hermes Agent — https://github.com/NousResearch/hermes-agent — MIT. Nothing of the runtime is redistributed here.

Memory: Hindsight (recommended, optional, degrades gracefully). Hindsight gives an agent durable notes between sessions. Without it the pack works within a session only — nothing in these skills depends on it. The memory provider ships with Hermes Agent (plugin plugins/memory/hindsight); enable it with hermes memory setup (choose hindsight) and put HINDSIGHT_API_KEY in the profile's .env. Upstream: https://github.com/vectorize-io/hindsight — MIT; docs at https://hindsight.vectorize.io. The template ships the memory block commented out, so a fresh install never points at a provider that is not there yet. There is no separate Hermes-side Hindsight plugin repository: the commonly cited github.com/NousResearch/hermes-plugin-hindsight link is dead (HTTP 404, measured 2026-10-05), and the NousResearch organisation publishes other hermes-plugin-* repositories but no hindsight one. The provider bundled with Hermes Agent is the supported path, and it is not redistributed here.

Reporting

The pack ships the agent-reporting skill so the agent can hand its owner a standard ARP v1.2.1 report on demand — what it did, what is stuck, what it recommends. Reports describe the agent's own work from its own real sources; no metrics are invented and no person's data is described. It is additive: it changes nothing about how the other skills behave.

Licence

MIT — Copyright (c) 2026 TalentDelta — except skills/agent-reporting/, which is CC-BY-4.0 (© AgentMinds) and is redistributed under those terms with attribution. See LICENSE for both notices in full.

The one rule this pack is built around

Declare what you removed, and verify what you ship. A pack that quietly carries a private path, a dead link, or a credential-shaped string is worse than no pack — it moves the problem to whoever installs it. Scrub, declare, and re-check the external links at the moment you build.