Hermes Agent

[andreasdueren]

@creative567145 Thanks, that log was enough to identify it. This is a different issue from the earlier gateway status detection bug.

The fresh install was hanging before supervisor started because start.sh ran:

npx agent-browser install

without forcing non-interactive mode. On a new /app/data volume, npx prompts with “Ok to proceed? (y)”, but Cloudron has no TTY there, so startup never reaches supervisor and port 8000 never binds.

I pushed package version 0.4.32 with the fix:

npm_config_yes=true npx --yes agent-browser install

So fresh installs should no longer block at that prompt. Existing installs where the browser runtime is already seeded were likely unaffected, which is why this showed up on a clean installation.

[enerch]

This post is deleted!

[chrisbol]

great that hermes agent is now available on Cloudron. But I have some issues with dashboard.
With help of my local AI i have drafter this message...

OPEN QUESTION TO THE MAINTAINER (setting the scene)
Is the browser-side dashboard (including the Chat and System pages) an intended
surface for this Cloudron package? Or is the dashboard primarily meant for Kanban

• the /v1 API, with chat/sessions hosted on the local desktop/CLI and via the
  gateway's platform channels? This determines whether the report below is a bug you
  want to fix or an intentional scope decision.

If web pages ARE in scope: the Hermes codebase ships full hosted-web-chat support
and first-class subpath-reverse-proxy support (X-Forwarded-Prefix handling and
/assets rewriting in _serve_index), so the failure below looks like a small,
easy-to-ship nginx routing gap in the package rather than deliberate design. The
proposed one-block fix resolves every affected page at once.

The bug:
I run the Hermes Agent package on Cloudron. The dashboard shell loads fine;
/kanban, the left nav, and the Files page all work. But every dynamically-loaded
page is blank: clicking "Chat" and clicking "System" both give a completely empty
screen, with the browser console showing 403 (Forbidden, net::ERR_ABORTED) on
their lazy-loaded asset chunks:

/assets/ChatPage-*.js
/assets/SystemPage-*.js
/assets/xterm-BrP-ENHg.css  and  /assets/xterm-*.js
/assets/themes-*.js
/assets/useProfileScope-*.js
/assets/chat-activation-*.js
/assets/ModelReloadConfirm-*.js

ROOT CAUSE (confirmed, not speculation)
The package's nginx only routes /kanban/* to the dashboard backend
(127.0.0.1:9119), rewriting the prefix and setting X-Forwarded-Prefix: /kanban.
Hermes's _serve_index then rewrites the statically-referenced asset URLs in the
HTML to /kanban/assets/... -- which is exactly why the shell, the Files page, and
the Kanban page all work.

But pages like Chat and System are loaded at runtime via dynamic import() with
build-time absolute paths (/assets/...). The browser therefore requests their
chunks WITHOUT the /kanban prefix. Those requests land on 'location /' -> the
front landing page's try_files -> fall through to the headless API server (port
8642), which has no SPA assets and returns 403/404. The dashboard backend itself
serves every chunk fine (returns 200), so this is purely a reverse-proxy routing
gap in the package, not a Hermes bug.

Quick repro that isolates it (from a Cloudron terminal):

curl -s -o /dev/null -w "%{http_code}\n" http://127.0.0.1:8000/kanban/assets/ChatPage-*.js   # 200
curl -s -o /dev/null -w "%{http_code}\n" http://127.0.0.1:8000/assets/ChatPage-*.js          # 403/403

PROPOSED FIX (in the package's nginx.conf)
Add a location /assets/ block that proxies to the same dashboard backend with the
same headers as /kanban/, including X-Forwarded-Prefix: /kanban so the CSS route
keeps rewriting url(/fonts/...) references correctly:

location /assets/ {
    proxy_pass http://127.0.0.1:9119;
    proxy_http_version 1.1;
    proxy_set_header Host 127.0.0.1:9119;
    proxy_set_header X-Real-IP $remote_addr;
    proxy_set_header X-Forwarded-For $proxy_add_x_forwarded_for;
    proxy_set_header X-Forwarded-Proto $scheme;
    proxy_set_header X-Forwarded-Host $host;
    proxy_set_header X-Forwarded-Prefix /kanban;
    proxy_read_timeout 600s;
    proxy_send_timeout 600s;
}

This single block fixes every lazy-loaded page at once (Chat, System, and any
future dynamically-imported route).

Environment: Hermes Agent v0.20.0 (2026.8.3), Cloudron deployment, dashboard bound
to 127.0.0.1:9119, SPA mounted under /kanban.

[andreasdueren]

@chrisbol Thanks for the detailed report, the root asset routing issue was confirmed and fixed in package version 0.4.65. The package now proxies the lazy loaded dashboard assets to the Hermes dashboard backend while keeping them protected by the same authentication as the Kanban dashboard. Chat, System, terminal, theme, and related asset chunks now return successfully. The update is published and available.

[msbt]

@andreasdueren do you have plans or even already tried adding something like https://hermes-workspace.com/ to the package?

[andreasdueren]

@msbt I’ll look into it

[andreasdueren]

@msbt this looks like a community package and probably should not be included on the Hermes package. Adds extra complexity, package churn and attack surface.

[msbt]

@andreasdueren aye, thanks for checking it out!

[alpheus]

@andreasdueren hello Andreas!
I have an issue with setting the HERMES_REDACT_SECRETS for scheduled cron jobs. Scheduled cron does phone number extraction and reports it to telegram, but the phone numbers arrive redacted (+49*****1234). If I create a one-off cron and then execute it via hermes cron ... then the redaction doesn't happen (phone numbers are visible).

I think the issue is that the main gateway process doesn't pick up the value that I wrote in the .env file (false), which results in phone numbers being redacted when they are outputted to telegram.
If you want to go into the issue with much more depth, below is the detailed summary from my agent of all the things we tried and all the findings that might be useful.

This happened on HermesAgent app version 0.4.61, but I would guess that the issue still persists.
Thank you for your time and dedication to the project maintenance!

Agent summary

Short version
python -m gateway.run imports agent.redact before gateway.run loads ~/.hermes/.env, so HERMES_REDACT_SECRETS=false arrives too late for its import-time snapshot.
Fixing that alone won't solve it though: the gateway's final outbound delivery call forces redaction regardless, and scheduled cron output passes through both scheduler and gateway redaction layers.

Root causes identified

• agent/redact.py snapshots HERMES_REDACT_SECRETS at import time (_REDACT_ENABLED = os.getenv(...)), so it must be set before agent.redact is imported.
• The Cloudron supervisor's environment= list for hermes-gateway includes HERMES_HOME etc., but not HERMES_REDACT_SECRETS/HERMES_REDACT_PHONES.
• gateway/run.py imports agent modules (which import agent.redact) near the top, but only calls load_hermes_dotenv() much later (~line 1829) — so the .env loads after the snapshot is taken.
• gateway/run.py::_redact_gateway_user_facing_secrets() calls redact_sensitive_text(text, force=True) — this bypasses _REDACT_ENABLED entirely, so security.redact_secrets: false in config.yaml has no effect on this path.
• The phone regex also misses Serbian local format and international numbers with a space after the country code.
• cron/scheduler.py redacts no_agent stdout/stderr independently, then the gateway applies its own forced pass — so scheduled jobs go through two redaction layers.

What we tried

• We tested: security.redact_secrets: false in config.yaml
• HERMES_REDACT_SECRETS=false in .env
• editing supervisord's environment directly (not durable — file is app-owned and replaced on updates)
• a sitecustomize.py early-patch (confirmed _REDACT_ENABLED=False standalone, but scheduled Telegram delivery still redacted due to force=True)
• disabling phone matching + HERMES_REDACT_PHONES=false (still redacted in the real delivery path)

[andreasdueren]

@alpheus I pushed 0.4.87 which should resolve this

[WiseMetalhead]

Am I right in thinking that Incoming Email in the Cloudron app settings currently doesn't work? Whenever I try to configure the mailbox, I always get these errors in the logs:

Sep 02 17:03:52 2026-09-02 14:03:52,991 ERROR hermes_plugins.email_platform.adapter: [Email] SMTP connection failed: STARTTLS extension not supported by server.

Sep 02 17:03:53 2026-09-02 14:03:53,013 WARNING gateway.run: ✗ email failed to connect

And if I send an email to that address, the agent doesn't respond at all.

Outgoing email works fine, but only if I leave the default settings unchanged after installing the app.

[alpheus]

@andreasdueren thank you for the fix! I will update hermes in the next few days to the latest version and let you know if it fixed it! Thanks for the fast effort and update!

[andreasdueren]

@WiseMetalhead Please update to 0.4.88. It should work now.

[nostrdev]

@andreasdueren we were initially tempted to use cloudron for packaging agents but given the many documented cases of AI breaking out of containers, compounded by the amount of unencrypted internal cloudron comms and ports being available, we backed away

We are now evaluating firecracker on a dedicated machine for agent hosting

Curious what your (and others) position is with regard to the risks of running agents on cloudron containers

[timconsidine]

@nostrdev E2B ? Kata ? Something else ?
(Probably a slippery slope needing its own discuss or off-topic thread, based on your thought-provoking question touching many surfaces).

[nostrdev]

slippery and off-topic indeed - put some iterations into it and made a new topic here. it's a subject that many, if not most, of us are interested in, has probably been well discussed before.

https://forum.cloudron.io/topic/15921/running-ai-agents-on-cloudron-infiltrating-other-containers-and-admin-services

[alpheus]

@andreasdueren Hi again!
I updated to the latest version (0.4.89) and I can confirm that HERMES_REDACT_SECRETS is successfully read from scheduled cron jobs!
Thanks for the fix!

Now I have a new but similar issue issue with GITHUB_TOKEN - it is present in the .env (agent confirms that in the chat), but it's not being read in the cron.

Here is the report from the agent:
The cron scheduler runs inside the gateway process. When it spawns the no_agent script, it builds the child env via
build_subprocess_env() → _sanitize_subprocess_env()
which unconditionally strips GITHUB_TOKEN / GH_TOKEN (they're in _ALWAYS_STRIP_KEYS in tools/environments/local.py).

The token is present in /app/data/.hermes/.env (line 436, uncommented, len 93) and in my interactive shell, but it never reaches the cron script's subprocess. This is a design conflict: the backup runner needs the token, but Hermes' secret-scrub policy deliberately removes it from child processes. GITHUB_TOKEN is not registered in config.yaml terminal.env_passthrough, so the scrub isn't bypassed.

Proposed fix (from our side, not yours)
A small wrapper-only code change in our python script that uses github for backup – not a config.yaml change and not a gateway restart.

Current flow:

cron scheduler (sanitized env; token stripped)
  → github_backup_publish.py
      → run_github_backup.py --publish

Proposed flow:

cron scheduler (sanitized env)
→ github_backup_publish.py
    → read /app/data/.hermes/.env internally
    → extract only GITHUB_TOKEN
    → start run_github_backup.py with the existing sanitized env + only GITHUB_TOKEN

After I did this workaround, the github token is successfully read.

My question for you

is this something you could also fix, or I should keep this (or some better) workaround?

Thank you once again for your time!

[andreasdueren]

I looked into this further. The missing token is a consequence of upstream Hermes deliberately filtering credentials from subprocess environments, rather than a Cloudron-specific environment issue.

Upstream does offer general environment passthrough, but it explicitly prevents that mechanism from overriding its protected credential blocklist. It also does not currently provide an operator-controlled, per-script credential allowlist for cron jobs.

I’m trying to keep this package as close to upstream as possible, so I won’t be adding a package-specific workaround for that filtering. Maintaining a separate credential-passing mechanism would add complexity and behavior that could diverge from upstream’s security decisions.

A scoped way to give trusted cron scripts access to selected credentials seems like a reasonable upstream feature request. I’d prefer to use an upstream-supported solution if one becomes available, rather than implement a parallel mechanism in the Cloudron package.

@alpheus

[alpheus]

@andreasdueren thank you for the response! Makes sense, I will use the current solution until hermes releases the feature.

Once again, thanks for maintenance of the cloudron app!

[micmc]

I've installed hermes recently and the only issue I have as far as I can see is that the Chat in the Web UI does not seem to work it keeps disconnecting with this message at the top "Chat connection interrupted (code 1006). Reconnecting..." and "WebSocket connection failed".
On the other hand it works well in the app's terminal. So i wonder if this is on purpose for the sake of running in CR?
I've tried a few things even asked hermes how to fix it, but what it suggested did not work. And this is the main line of error message I get in the log

2026-09-15 14:52:01,841 WARNING hermes_cli.web_server: pty refused: origin_mismatch origin=https://hermes-agent.xxxx.xxx bound=127.0.0.1 peer=127.0.0.1

I've looking in this thread but found not the same probem.
Anyone having the same issue or know how to fix it (if fixable in Cloudron of course)