Nerdit for Hermes Agent
Hermes puts what it builds online on its own machine, and keeps it alive.
Ask Hermes for an app, a bot or a scheduled job. It writes the code, then this plugin hands the
folder to Nerdit, the open-source engine running on the
same machine: Nerdit builds it, runs it as a supervised service and gives it a URL. The app
keeps running after the chat ends. When one of your apps goes down, Hermes knows on its next
turn, without being asked.
You (Telegram): build me a page that shows my server's uptime, and put it online
Hermes: ...writes the app, calls nerdit_ship...
It's live: https://nerdit.local/uptime/
You: /apps
Hermes: 🟢 uptime running https://nerdit.local/uptime/
🔴 feed-bot failed
What the plugin adds
nerdit_ship | One tool call that deploys a folder, waits until it is healthy and returns the URL. On failure it returns the engine's diagnosis, so Hermes fixes the cause and ships again. |
nerdit_apps, nerdit_diagnose, nerdit_logs | What runs here, why an app is failing, and its logs. |
nerdit_share | A hosted https URL for an app, through the machine's link to your Nerdit account. Private by default (you and the people you allow in the Nerdit App), public only after you confirm, and off to stop. |
nerdit_call | Every other tool of the engine by name: databases, local models, schedules, webhooks, sharing, and the engine's own guides (get_guide). Anything that is not a read or a restart asks you first. |
Outage awareness (pre_llm_call hook) | At the start of a session Hermes is told what runs on the machine. Later it is told once when an app becomes failed, degraded or restarting. No polling by the model, no tokens spent while everything is healthy. |
| Standing instruction (system prompt section) | Six lines that tell Hermes to ship through nerdit_ship and not to drive Docker or the nerdit CLI from the terminal. |
Approval guard (pre_tool_call hook) | Through nerdit_call, only reading (list_services, get_service, service_logs, diagnose_service, list_guides, get_guide, wait_for_service) and restart_service run directly. Every other engine tool asks a person first, whatever the session's approval mode, and so does any tool a later engine version adds. nerdit_share asks before making an app public or stopping a share. |
/apps | The status of your apps in the CLI, Telegram or Discord, answered by the plugin itself with no model call. /apps <name> shows why that app is failing. |
| "Keep my apps alive" blueprint | /blueprint nerdit:watch schedules a check every 5, 15 or 30 minutes. When an app is down, Hermes diagnoses it, restarts it if that is the fix, and reports. It stays silent while everything runs. |
| Skills | nerdit:ship and nerdit:rescue, the two workflows above written out. |
Install
You need the Nerdit engine and Docker on the machine where Hermes runs. Follow the
engine quickstart, then check it with nerdit doctor.
hermes plugins install nerdit # or, before it is listed: hermes plugins install nerdit-ai/nerdit-hermes
hermes plugins enable nerdit
Run nerdit trust once so that the https://nerdit.local/... URLs open without a certificate
warning. Start a new session and ask: "build a hello-world web app and put it online".
If Hermes runs as a service with a minimal PATH, set NERDIT_BIN to the full path of the
nerdit executable. Without the engine the plugin stays quiet: its tools answer with a clear
"not installed" error and the hooks add nothing.
What this plugin does on your machine
- It starts a local process. Each call runs
nerdit mcp, the server that ships with the
engine, and sends it one request over stdio. No shell is involved and the arguments are
fixed. The outage check runs it at most once a minute, with an 8 second limit. - That process does not inherit Hermes's environment. It receives only
PATH,HOME,
USER,LANG,LC_*,TMPDIR,DOCKER_HOST,DOCKER_CONTEXTandNERDIT_*. The
provider keys and tokens Hermes holds never reach the engine. If your engine needs another
variable, set it where the engine's daemon starts, not in Hermes. - It acts with your
nerditCLI's rights. The plugin never reads, stores or sends a
credential. Thenerditprocess uses the token your CLI is configured with, which is an
admin token by default. To narrow what Hermes can do, give the CLI a scoped token, as
described in the engine's security guide. - It reads the folder you ask it to deploy. The engine zips that folder and builds it. The
plugin reads nothing else outside its own files. - Network. The plugin makes no network call. The engine listens on
127.0.0.1. Apps you
deploy do whatever their code does. A URL other people can open (share_service) goes
through Nerdit Cloud and needs a linked account, which is optional. - Secrets. The skills tell Hermes never to ask for a secret in chat. You set them yourself
withnerdit secrets set <app> --prompt KEY. - No telemetry, no background thread, no self-update.
- Unattended runs. The scheduled watch is instructed to read, diagnose and restart only,
never to delete or redeploy.
Develop
python3 tests/test_plugin.py # stdlib only, uses a fake engine
hermes plugins validate . --install-deps
Apache-2.0. Engine: https://github.com/nerdit-ai/nerdit · Docs: https://docs.nerdit.ai
README from the reviewed commit ca6e576; it updates when the author re-pins.

Nerdit