Agent on Demand Hermes Template

This is the first quick Tenki template for the Agent on Demand workstream. It starts from Tenki's sandbox base image, installs the Hermes Agent command-line harness, and includes runtime scripts that configure Nebius Token Factory as the backing OpenAI-compatible provider.

The template does not bake API keys into the image. Run the runtime setup inside each sandbox session with NEBIUS_API_KEY loaded.

Why This Shape

Tenki's current template system is Git-backed: .tenki/template.json points at a Git repository and ref, then Tenki builds the image from that context. Local archive contexts are not supported in the template docs checked on 2026-08-04.

The Tenki base image already includes Ubuntu 24.04, Node 24, npm, pnpm, Bun, Python 3.12, git, curl, jq, rg, build tools, and terminal coding agents. The build step only installs the missing learner-specific layer:

• Hermes Agent
• convenience setup commands in ~/.local/bin
• starter workspace material

The agent harness is Hermes Agent. NousResearch/Hermes-4-70B is only one optional backing model exposed through the Nebius provider alias nebius-hermes-model.

Build

Local Docker validation before Tenki:

npm run validate
npm run check:shell
npm run docker:build
npm run docker:check

Docker is only a local parity harness. Current Tenki Sandbox template docs use Git-backed .tenki/template.json builds, not local Docker image imports. See docs/local-docker.md and docs/docker-validation-guide.md.

For closer Tenki-style local operations:

node scripts/local-tenki.mjs template build agent-on-demand-hermes
node scripts/local-tenki.mjs sandbox create --name agent-on-demand-hermes
node scripts/local-tenki.mjs sandbox exec --session agent-on-demand-hermes -- hermes --version
node scripts/local-tenki.mjs sandbox terminate --session agent-on-demand-hermes

The local build installs the core Hermes Agent harness, skips bundled Chromium, and time-boxes noncritical browser/TUI npm dependencies so the image build stays deterministic. Browser tools, ffmpeg, and deployment CLIs should be added as explicit modules once the core chat/session path is green.

The core build pins the Hermes Agent checkout to:

v2026.8.3

Latest local Docker validation resolved that tag to Hermes Agent v0.20.0 at commit 3c27eb6234bf91b8ceee9e9071591b31e9b148cb.

Update .tenki/template.json once this folder lives in its final GitHub repo:

tenki template init agent-on-demand-hermes \
  --repo https://github.com/fruteroclub/agent-on-demand-hermes-template \
  --ref main \
  --workdir /home/tenki/agent-on-demand-hermes \
  --force

Then build:

tenki template build agent-on-demand-hermes

The build prints the image ref. Launch a session from that image:

tenki sandbox create --name agent-on-demand-hermes --image agent-on-demand-hermes

Configure A Session

Inside the session, load the key without printing it:

read -rs NEBIUS_API_KEY
export NEBIUS_API_KEY
agent-on-demand-setup

Validate the non-secret config:

agent-on-demand-check

Run a tiny paid smoke test only when you are ready to spend a request:

hermes

Then ask:

Reply exactly: hermes-template-ok

Learner Prompt

Start the harness:

hermes

Use hermes --tui only after the optional TUI npm dependencies have installed successfully.

Then ask:

Create a small static website in ./site for an AI builder workshop. Use plain
HTML, CSS, and JavaScript. Include a landing section, a three-step agenda, and a
small interactive checklist. Add an npm script that serves the site locally.

Cleanup

When the tutorial is done:

tenki sandbox terminate --session agent-on-demand-hermes

Current Live Checks

Checked from this machine on 2026-08-04:

• Tenki CLI: v1.0.0
• Local Tenki auth: invalid/revoked, so remote template build was not run
• Hermes Agent quickstart/install docs checked
• Token Factory catalog included Qwen, Hermes-4, Nemotron, and DeepSeek models
• Tenki docs base image page was last updated 2026-08-03
• Tenki docs template page was last updated 2026-08-03