A trustable, open-source dev harness for people building real businesses with AI agents.
A local guardian agent (Anjo) that interrogates you on first run and orchestrates your work, paired with a 24/7 remote orchestrator (Hermes) that runs on any VPS and handles alarms, communication, and cron — with clear steps, logs, visibility, security, plans, and reviews baked in.
Most "agent setups" are a pile of personal scripts, secrets, and tribal knowledge that only the
author can run. Anjo Harness is the opposite: a clean, documented, reproducible template you can
clone, run ./install.sh, answer a few questions, and have a working dev environment plus an
always-on orchestrator you actually trust.
No real data ships in this repo. Everything personal is asked at setup time and written to git-ignored config. You stay in control.
| Component | Where it runs | What it does |
|---|---|---|
| 🪽 Anjo | Your machine (local-first) | First-run interrogation wizard, config generation, local orchestration, dispatching plans, talking to Hermes. |
| 🐎 Hermes | Any VPS (Hetzner reference) | 24/7 orchestrator: alarms, communication channels (Telegram/email/webhook), cron, health, activity stream. |
Anjo asks. Hermes never sleeps. Together they give you a harness with visibility, security, and well-elaborated plans → implementations → reviews.
git clone https://github.com/YOUR_ORG/anjo-harness.git
cd anjo-harness
./install.sh # bootstraps deps, then runs the Anjo interrogation wizardOn first run Anjo will interrogate you in the terminal:
- Project name, language, and workspace layout
- Whether you want a remote Hermes (and on which VPS)
- Which communication channels to wire for alarms (Telegram, email, webhook)
- Secrets handling (sops + age, or plain git-ignored
.env)
Everything you answer is written to git-ignored local config. Nothing personal is committed.
anjo status # health of local + remote
anjo plan new "ship feature X" # scaffold a plan from the unified template
anjo orchestrate # dispatch the active plan, stream progress
hermes deploy # provision/refresh the remote orchestrator- One planning template per layer — every directory has a single
.planning/TEMPLATE.mdthat consolidates plans, specs, and status into one well-structured doc that the orchestration refines. - Reproducible infra — Hetzner cloud-init, Docker Compose, Caddy TLS, systemd units, sops+age.
- Alarms & comms — Hermes pushes every alarm to the channels you chose, with an activity stream you can tail live.
- Plans → implementation → review loop — the harness enforces the cycle, not just the code.
- Security first — no secrets in git, encrypted env, least-privilege, documented threat model.
anjo-harness/
├── install.sh # one-command bootstrap → runs the Anjo wizard
├── anjo/ # local guardian CLI (Python stdlib, zero-dep)
│ └── .planning/TEMPLATE.md
├── hermes/ # remote orchestrator (Docker Compose, alarms, comms)
│ └── .planning/TEMPLATE.md
├── infra/ # Hetzner / VPS setup, sops, firewall
│ └── .planning/TEMPLATE.md
├── docs/ # PRINCIPLES, OBSERVABILITY, SECURITY, master planning template
└── .github/ # CI, issue/PR templates
make install # create .venv and install dev tooling (pytest, ruff)
make test # run the test suite
make lint # ruff
make smoke # zero-dep end-to-end smoke (wizard + orchestrate + imports)
make help # list all targetsThe runtime (Anjo + Hermes) is pure standard library — only the dev tooling needs installing.
v0.1.0 — scaffold. See CHANGELOG.md and the per-layer .planning/TEMPLATE.md
for what's planned next.
MIT — use it, fork it, build your business on it.