Dashboard or Webapp: choose the right browser surface
The Hermes Dashboard and the gateway Webapp are different browser interfaces. Use hermes dashboard for the administration panel, normally on port 9119. Use hermes webapp for the Desktop-style workspace served by the messaging gateway, normally on port 9120. Both are self-hosted; opening either in a browser does not transfer responsibility for your computer or VPS to a hosting service.
The Webapp path below is documented in the official v0.23.1 source and current Desktop browser-workspace guide, checked November 16, 2026. Older installations and packaged builds can differ. Record the version on the machine running the gateway before following new instructions; a newer laptop client does not update a remote host.
Choose by the job, not by the word Web UI
- Administer an installation: Dashboard provides configuration, credentials, sessions, logs, skills, MCP and operational controls. Its Chat pane embeds the TUI over PTY/WebSocket; it is not a chat-free monitoring screen. Follow the Dashboard setup walkthrough.
- Work in the Desktop-style interface from a browser: Webapp provides sessions, profiles, files, Git review, artifacts, previews and a terminal rail through the owning gateway. Native-only features such as global shortcuts, OS-window inspection and the updater remain in Electron. Browser support is not parity with every native feature.
- Connect the native Desktop application to another host: use the remote-backend decision guide. A Desktop Remote gateway URL, a Dashboard URL and a Webapp URL are not interchangeable just because they reach the same machine.
- Avoid running the host yourself: compare managed hosting and allowances. FlyHermes is independent of Nous Research. Self-hosted browser access already exists; the commercial distinction is who operates the runtime, updates, authentication and recovery, not who is allowed to chat in a browser.
Open the gateway Webapp without exposing it
These are documented setup steps, not a live-agent test performed by this site. First finish active work and identify the service account and existing gateway owner. The restart below is a real runtime change; do not run it casually against another person's active agent.
# Run on the host that owns the intended Hermes installation.
hermes --version
hermes config set webapp.enabled true
# After active work finishes, using the authorized gateway owner:
hermes gateway restart
hermes webapp --status
hermes webapp
The default address is http://127.0.0.1:9120. Use the URL printed by your installation rather than forcing that port: standalone profile gateways can use another saved port. For a named profile, the documented entry point is hermes -p work webapp, replacing work with an existing profile. See the profile and state-boundary guide.
For local access without a sign-in gate, the launcher creates a private one-use link valid for 60 seconds. Treat an unused link as a credential; do not paste it into a support issue, screenshot or shared chat. A fresh tab or bare address can show sign-in instructions rather than your workspace. Run hermes webapp again instead of copying stale links. The current guide documents hermes webapp --sign-out as signing out Webapp tabs, not every separately authenticated native client.
Secure remote access on the correct listener
The Webapp configuration reference separates webapp.host, webapp.port and webapp.public_url from the standalone Dashboard's address. Setting dashboard.public_url does not publish the Webapp. A same-machine reverse proxy can forward to the loopback Webapp while webapp.public_url names the external HTTPS address; this requires sign-in even when the listener stays on loopback. A proxy in another container cannot reach that container's loopback listener.
Prefer a private tunnel or VPN. For a trusted LAN/VPN, the documented shared-password provider is an option; use OAuth/OIDC for an internet-facing deployment. Do not router-forward port 9120 as a shortcut. A signed-in Webapp user can run commands, access host files and reach the Hermes OS account's environment and credentials. If saved gateway connections are available, anyone who can sign in can also use those connections as their saved identities. Review that inventory before sharing access.
Plain HTTP on a non-loopback address also loses browser secure-context features: microphone, clipboard reading, notifications and the bot screen viewer can be unavailable. Use HTTPS or a tunnel opened as localhost; do not diagnose every missing button as an agent or model failure. The phone access guide covers the client choice separately from hosting.
Verify the workspace, not only a green status
- Check
hermes webapp --statuson the owning host. If the port is taken, authentication is required but unconfigured, or the browser sign-in file is inaccessible, the gateway may still serve native clients while Webapp remains unavailable. Fix the reported cause; don't reset memory or reinstall every surface. - Open a fresh launcher link or use the configured sign-in. Confirm the gateway, selected profile and a non-sensitive working folder before allowing tools to act.
- Complete one harmless text exchange, then separately authorize a read-only file check. Inspect the actual result; successful HTTP or a loaded transcript is not proof of a tool turn.
- If an older conversation is read-only, current docs distinguish adoptable pre-gateway chats from rows created by a separate WebUI agent and historical automation runs. Review
hermes sessions adoptwith the owning profile and a backup; do not promise it makes every historical run writable.
An isolated dashboard is not tenant isolation
hermes dashboard --isolated gives a named profile a dedicated server and a narrower visible roster. It does not sandbox sibling profiles. Current Dashboard profile documentation explicitly warns that profiles under one OS user can reach each other's files through filesystem routes or agent tools. Different passwords on those servers do not change the filesystem boundary.
For mutually untrusted people, use separate OS users or containers with appropriate filesystem and network separation. Before buying managed hosting for a team, verify its actual access policy rather than assuming a profile switcher means tenant isolation. Keep the server security checklist and backup procedure in the operating plan.
Sources checked November 16, 2026: official Desktop Use the Desktop workspace from a browser, Open it from your phone or another computer, Reach your other gateways from one Webapp; Configuration Webapp; Dashboard Managing multiple profiles. These establish documented mechanisms and boundaries, not measured uptime, a completed migration or a guarantee that an interrupted task ran exactly once.
Remote dashboard has no sign-in button? Check the authentication mode
A reachable Hermes dashboard is not always a browser-login deployment. Current official docs describe a deployment session-token provider as well as username/password, Nous OAuth and OIDC. A token-only deployment can be working as designed when its login page explains the credential channel instead of offering a sign-in button. Do not disable authentication or add a second password just to make that screen look familiar.
Choose by client and exposure
- Local browser: start on loopback with
hermes dashboard; use the Web UI setup guide to verify the printed address and selected profile. - Private remote Desktop: a deployment that already provisions
HERMES_DASHBOARD_SESSION_TOKENcan use Desktop's Session token field. The bundled provider registers when that credential is provisioned, unless explicitly disabled. This is a shared deployment credential, not a personal browser account. - Browser login on a trusted LAN or VPN: the built-in username/password provider is an option. It is not the recommended public-internet login path.
- Shared or internet-facing access: prefer Nous OAuth or a conformant OIDC provider. A deployment token has no per-user identity. Keep transport and origin access protected; knowing an admin token grants full dashboard access.
The official token-provider reference documents secret-manager provisioning and file:<path> loading at startup. Do not paste real credentials into a URL, public issue, screenshot, shell history or article example. A secret file must not be placed in the world-readable Nix store. Follow the server security checklist for the deployment's exposure boundary.
Read status without treating it as a successful login
Run this read-only check only against your own backend, over its established private or protected connection. Replace the example URL; the example contains no credential:
DASHBOARD_URL='http://127.0.0.1:9119'
curl --fail --silent --show-error "$DASHBOARD_URL/api/status" \
| python3 -m json.tool
Inspect auth_required and auth_providers. In the documented token-gated case, the gate is on and the provider list includes session-token; other configured providers may also appear. The status endpoint is public, so a 200 response proves reachability, not authorization. An authorized /api/auth/me request and a real Chat connection are separate acceptance checks. Do not print a token to construct a diagnostic log.
After token rotation, update the saved client credential
The official reference warns that rotating the deployment secret invalidates saved client copies. This can look like HTTP reachable but WebSocket fails in Desktop. Confirm that the client points to the intended backend, have the authorized operator update its Session token field through the secure client interface, then reconnect. Do not rotate unrelated model-provider keys or clear session history to repair a stale dashboard credential.
The provider verifies both REST bearer authentication and the token-mode WebSocket credential. This does not make every connection error a token error: check the backend's loaded provider, the intended host/profile and the same-window redacted logs. Use the remote Desktop diagnostics for authentication versus Host/peer rejection, and the profile guide before assuming a missing conversation was lost. For phone-browser access, the mobile guide explains why a Desktop token field is not a browser sign-in flow.
Decide who will maintain access
Self-hosted Web UI includes real browser Chat. You still own secret provisioning, client updates after rotation, protected networking, backups and recovery. This site promotes FlyHermes independently of Nous Research. Compare self-hosted versus managed responsibilities and current managed plan allowances if operating that lifecycle is the obstacle. Hosting does not establish external account permissions or guarantee recovery of an interrupted task.
Sources checked November 2, 2026: official Dashboard Deployment session-token provider, Authentication (gated mode) and Connecting Hermes Desktop to a remote backend. These are documentation-based decision and acceptance checks; we did not rotate a live token, deploy a token-only server or measure its reliability. No first-supported-release claim is made: check the running backend's documentation/version rather than assuming an older package includes this provider.
Fix a hub setup that cannot find the selected profile's credentials
A dashboard can open successfully while a skill or MCP setup fails to find its required credential. First check which profile the setup targets, not just which account launched the server. The current official profile-management documentation says hub actions—skill install, update and uninstall, MCP install, and toolset setup—use the target profile's .env and configured secret sources, rather than the dashboard process's environment. This also applies when the target is default.
Before installing or rotating a key
- Select the intended profile in the sidebar. Check the
?profile=URL value and, when shown, the amber managed-profile banner. Set as active on a profile card changes the sticky default for future CLI/gateway runs; it does not change the dashboard's selected write target. - Inspect that profile's API Keys page for the required key's presence. A key available to the shell that launched the dashboard is not proof that a hub setup can access it. Do not copy the entire default
.envinto a bot profile to make a setup pass. - From a trusted terminal on the same backend, confirm the target file locations without printing their contents:
hermes -p work config path
hermes -p work config env-path
Replace work with an existing profile. If the credential belongs to an external secret source, verify the target profile's configured source rather than adding a competing copy blindly. For profile boundaries and inherited environment caveats, use the profile guide.
Verify setup, availability and a real request separately
After an authorized credential correction, retry only the intended setup and inspect its result. For MCP, the dashboard Test action connects, lists tools and disconnects; this checks connectivity, not whether every tool can complete a task. Skill toggles take effect on the next session. Start a fresh Chat under the same profile and test one permitted, read-only action against non-sensitive data. A provider request may still incur charges.
If setup succeeds but the action fails, preserve the error category, profile, backend version and timestamp. Follow provider and quota diagnostics for billing/auth errors; do not widen permissions or expose credentials as a generic repair. The dashboard setup guide covers the launch and access checks. Compare managed hosting costs only if operating the host is the obstacle—hosting is not proof of external account permission.
Source checked October 26, 2026: official Dashboard sections Managing multiple profiles, Profiles, Skills and MCP. This is a documentation-based diagnostic checklist, not a reproduced credential failure or a promise that every missing-tool report has this cause.
Restore a dashboard chat that says it is open in another window
A visible transcript, a live browser connection and ownership of a running chat are different things. If New chat works but returning to an earlier conversation says “This chat is open in another Hermes window”, first establish whether another client really owns it. Do not delete the session database, remove lock files or kill every Hermes process to make the warning disappear.
Separate the failure before changing anything
- The page will not open: follow the localhost and service-owner checks above. No listening dashboard is a different failure from a refused session.
- Status loads but Chat cannot connect: check authentication and the WebSocket path. A public
/api/statusresponse does not prove that/api/ptyor/api/wsaccepted the client. The remote-backend guide separates those checks. - One saved chat is refused while a new chat works: record the selected profile and session ID, then inspect other browser tabs, Desktop windows and terminals on that backend. A genuine live owner should not be displaced by a second writer.
- Chat reconnects but the result is unclear: inspect the same conversation and the real output before retrying. A missing final answer is not evidence that a file write, message or deployment failed.
The documented orphaned-PTY case
Upstream fix #133847, merged October 6, describes a narrower failure: starting a new dashboard chat rotated the tab's attachment token, but the previous detached terminal process could retain the old session's single-writer lease. Returning to that conversation then produced the “another Hermes window” refusal even though the user had moved away from it.
The fix targets only the detached PTY holding the requested conversation, using the resumed session and its profile's lease holder. It deliberately leaves a terminal with a live viewer and detached terminals for other sessions alone. This is not a reason to bypass normal session ownership.
The published v0.22.2 tag contains that merged fix; this was checked against upstream release ancestry on October 19, 2026. That does not establish the first release containing it or prove that your running backend has updated. Record the version shown by the actual dashboard, not only a newer CLI installed on your laptop. Follow the supported update guide, preserve the owning profile and let active work finish before restarting the intended backend.
There is an important limit: the fix releases the stranded terminal; it does not promise to preserve a turn still running inside that terminal. The upstream change explicitly warns that closing it interrupts in-flight work. Inspect completed outputs before requesting the remaining work again.
A safe restore acceptance check
Use a non-sensitive workspace and an inexpensive model. These are proposed checks for your installation, not a benchmark we ran against your agent.
- In Chat A, ask for a short text-only reply with a distinctive harmless marker. Record the profile and session ID; wait for the reply to finish.
- Start Chat B and complete another text-only exchange. Return to A through the session list or Sessions → Resume.
- Confirm A's marker and history, then send one new text-only message. Passing means the intended conversation accepts it without the unexpected ownership refusal; a green Status page alone is insufficient.
- Separately test a temporary network interruption or hiding and returning to the tab. Check that you returned to the same session and did not submit the instruction twice. Do not combine this with a backend restart and call it the same test.
If it still fails, retain the time, backend version, profile, session ID, exact error and redacted logs. Note whether the sequence was A → B → A, a network drop or a process restart. Never include cookies, access tokens or private transcripts in a public issue. The setup walkthrough covers the initial Chat check; the phone guide covers client suspension.
Choose who owns recovery
Disclosure: this site promotes FlyHermes, a managed service independent of Nous Research. Self-hosted Web UI includes real TUI-backed browser Chat. You remain responsible for the host, authentication, updates, backups and incident recovery. FlyHermes is the managed alternative when you want less operational responsibility, not a prerequisite for browser chat and not proof that every interrupted task completes exactly once. Compare self-hosted and managed responsibilities and verify the hosted service's supported workflow before moving consequential work.
Sources checked October 19, 2026: Dashboard Chat and remote-backend reference, TUI connection recovery, targeted lease fix, and v0.22.2 release. A socket reconnect, a resumed transcript and recovery after process death have different guarantees.
Choose the next check by the failed control
If the dashboard is not running yet, follow the Web UI setup sequence before configuring channels. If Save succeeds but Restart fails, use gateway restart troubleshooting: the account that can edit a form is not necessarily authorized to restart a system service.
For a remote host, record the supervisor and account in the VPS operator checklist. A laptop that connects successfully through Desktop remote-backend setup has proved a client connection, not messaging delivery or permission to administer the server.
After a change, use the monitoring control-state checks to distinguish an intentionally disabled webhook from a restart-dependent channel change. Compare self-hosted and managed responsibilities if nobody owns this maintenance, then review managed plan allowances against the actual workflow. Managed hosting does not make external account permissions or usage limits disappear.
Source checked October 16, 2026: the official Channels controls specify saved credentials, enablement, connection testing and gateway restart requirements. These are documented operating procedures, not a claim that this publication run connected a bot or restarted a user's service.
Session cleanup refused: protect live work before pruning
A dashboard that can display sessions is not necessarily allowed to delete them. Current official session-maintenance documentation, checked October 31, 2026, distinguishes a guarded single-session deletion from offline bulk maintenance. Treat a refusal as a safety boundary, not a reason to delete the database or retry in a loop. These are documented procedures and proposed acceptance checks, not a destructive test we ran on a production gateway.
Read the refusal before choosing recovery
- client_upgrade_required: the delete request has no mutation identity. The official reference names Desktop 0.22.2 and older as affected clients. Update the client that issued the request; a newer backend alone does not upgrade that client.
- mutation_identity_required: the request carries only part of the required identity. API clients must read the mutation snapshot and send its request ID, expected revision and expected generation. Do not invent values or bypass the revision fence.
- session_busy: a turn is running or queued. Wait for it to finish and inspect its actual output before trying to remove its history.
- unknown_execution: a prior turn was lost across a gateway restart. Establish whether its external action happened before attempting recovery. An absent final reply does not establish failure.
- Exclusive maintenance refused: bulk delete, empty-session deletion, non-dry-run prune and owner backfill need the profile's exclusive gateway lock. The current documentation says a live gateway causes HTTP 409 and no change. Read endpoints and dry-run prune remain available.
These are different HTTP 409 cases. None is evidence that the session database is corrupt. Use the memory and session recovery guide only when diagnostics actually identify a storage problem.
Plan a profile-scoped maintenance window
- Identify the owning host and profile in the dashboard. Use the profile isolation guide to avoid stopping another bot's gateway. Record the intended session IDs, retention rule and rollback owner.
- Inspect active and queued work, arrange a quiet window and create a protected backup. A session export is an audit artifact, not proof of a restorable full backup. Do not publish transcripts or credentials.
- Preview the cleanup where a dry-run is available. Check the candidate list before approving irreversible deletion. A dry-run can be used while the gateway is live; it is not the actual cleanup.
- Drain work and stop the intended profile's gateway through its existing authorized service owner. For an explicitly named profile, the command pattern is shown below. If a container or supervisor owns it, use that deployment's established stop procedure; confirm it stays stopped rather than racing automatic restart.
- Run only the reviewed maintenance operation in that same profile, then restart through the same service owner. Inspect Sessions and Logs and verify the expected retained history and one permitted channel reply. Follow gateway recovery checks if restart or delivery fails.
# Replace work with the existing profile that owns the target history.
# Read-only orientation first; do not stop unrelated profiles.
hermes -p work gateway status
# Only after work is drained and a maintenance window is approved:
hermes -p work gateway stop
# Perform the reviewed cleanup in that profile's dashboard; no delete-all recipe.
hermes -p work gateway start
hermes -p work gateway status
Do not remove gateway.lock or SQLite sidecars to defeat the refusal. The lock protects a running owner; direct SQL deletion bypasses the documented mutation controls. Single-session deletion also retains small terminal-receipt tombstones indefinitely after executed turns, according to the delete endpoint reference. Deleting visible history therefore is not a promise of complete physical erasure or removal of every execution receipt.
Decide who will own cleanup and recovery
Self-hosted Web UI includes browser Chat; maintenance responsibility is the distinction. Someone must own retention choices, backups, the quiet window, service restart and delivery verification. This site promotes FlyHermes independently of Nous Research. Compare managed hosting and plan allowances when nobody wants to operate that lifecycle, but confirm retention/export policies and supported workflows before moving sensitive sessions. Managed hosting is not a guarantee of unlimited storage, total erasure or exactly-once task execution.



