Try it locally¶

A local Rocket.Chat, Hermes Agent and this plugin, wired together by Docker Compose. A bot, two people (alice and bob) and a room (#hermes-demo) are created for you. Everything listens on 127.0.0.1 only.

What you need

Docker with Compose. The first start downloads the Rocket.Chat, MongoDB and Hermes images, so it takes a few minutes.

Start¶

From a checkout of the repository:

git clone https://github.com/csabatatrai-ulx/hermes-rocketchat-team-gateway
cd hermes-rocketchat-team-gateway
docker compose -f try/compose.yaml up -d
python3 try/check.py

It starts with a scripted demo model: no account, no API key. It is not AI; it answers with fixed text, and asks to run a terminal command when you ask it to delete something. Everything around the model is real: the plugin, Hermes, Rocket.Chat, the command guard and the approval buttons.

try/check.py checks them end to end and exits non-zero if any check fails (Python 3, standard library only, about 15 seconds once the bot is up; on the first start it first waits a minute or two for the bot to connect):

PASS  a mention gets an answer that quotes the question and mentions the asker
PASS  two people at once: each answer quotes its own question and mentions its own asker
PASS  /yolo is refused over chat, without reaching the model
PASS  a dangerous command waits for approval, with buttons on the prompt
PASS  another person's approval runs nothing
PASS  the asker's approval runs the command
PASS  a direct message needs no mention

7 / 7 passed

Then choose a real model:

Nous Portal (free tier)Any OpenAI-compatible endpoint

# sign in to Nous Portal and pick a model; it replaces the demo model
docker compose -f try/compose.yaml exec -u hermes hermes hermes portal
docker compose -f try/compose.yaml restart hermes

Free-tier models can be slow: one took two minutes for a single reply in our test. meituan/longcat-2.5:free answered in 3 to 12 seconds, approval prompt included (tested 2026-11-11), so pick that one when asked for a model.

Start the stack with the endpoint instead (down -v first if it already ran):

LLM_API_KEY=... LLM_BASE_URL=https://api.example.com/v1 LLM_MODEL=model-name \
  docker compose -f try/compose.yaml up -d

The model needs tool calling for the approval demo.

Use¶

Open http://localhost:3000 and sign in as alice, password hermes-demo. Sign in as bob in a private window to be the second person.

Try this in `#hermes-demo`	What happens
`@hermes when is the release?` as alice and bob at once	Each gets their own answer, quoting the question
`@hermes please delete the old build` (demo model) or `@hermes run rm -rf /tmp/old-build with the terminal tool` (real model)	A prompt with buttons; only the asker's click runs it
`@hermes /yolo on`	Refused, with no model call
`@hermes !new`	A confirmation with buttons
A direct message to `hermes`	No mention needed

The terminal tool runs inside the Hermes container only.

Stop¶

docker compose -f try/compose.yaml down -v    # -v also deletes the demo data

Notes¶

Another port: TRY_PORT=3300 docker compose -f try/compose.yaml up -d, then open http://localhost:3300.
Your own changes: the plugin is mounted read-only from the checkout, so a change is live after docker compose -f try/compose.yaml restart hermes.
Not for production: the demo passwords and settings are for this local stack only. For a real server, see Install on your server.