A browser window does not tell you who operates the agent behind it. Hermes Desktop can connect to a server you maintain. A self-hosted Web Dashboard can open in your phone's browser. Neither turns that server into a managed service.
Quick answer#
Choose self-hosted Hermes Agent when you need control over local files, private-network tools, model routing or the execution environment, and you can own updates, security, backups and recovery. Choose FlyHermes when you want managed browser/mobile access, supported model access and connected channels without maintaining the host and gateway yourself. Check the current FlyHermes plans and usage terms before buying; managed hosting is not unlimited inference or a promise that every integration works without authorization.
The useful question is not “does it have a dashboard?” It is “where will my work run, and who fixes each part when I am away?”
Check what the sandbox actually contains#
A Docker terminal backend and a container running the whole Hermes agent protect different scopes. The official security policy, section 2.2, checked November 20, 2026, says terminal-backend isolation confines shell and file operations. It does not contain the agent process, code-execution child, MCP subprocesses, plugins or hooks. For untrusted inputs or shared production use, the policy calls for whole-process wrapping with intentional filesystem and network access.
Profiles are not tenants. Two profiles under one OS user can reach sibling files; a Dashboard profile switch or --isolated server does not change that. If different customers must not reach each other's state, use separate OS users or whole-agent containers with separate credentials and reviewed mounts. Do not share the Docker socket or mount a common secrets directory and call that tenant isolation.
Before buying managed hosting, ask which layer is isolated, who reviews installed extensions, and who controls account access, backups and exports. The public FlyHermes offer establishes hosted operation, not a customer-specific isolation design or contractual recovery guarantee. The self-hosted AI agent decision page now separates these questions from a generic feature scorecard. This is documentation-based guidance, not a security test or certification.
Decide the exit conditions before handing over the workflow#
Managed hosting should remove work you do not want to own, not obscure who can stop the bill or recover the output. Before subscribing, ask three concrete questions: what cancels renewal, what preserves useful state, and what refund rules apply? Keep the service name on the answers. A provider balance, stopped server and canceled managed plan are not equivalent states.
The FlyHermes refund policy says cancellation stops future renewals but does not automatically refund paid amounts. The first-account request window is 14 days; used AI credits are deducted and approved annual refunds include the policy's prorated unused months. Do not describe this as a free trial or guaranteed full refund. Mandatory consumer rights remain unaffected.
By contrast, the separate Nous Cloud FAQ says a stopped agent retains storage charges, while model and tool usage follows its current plan allowance and any pay-as-you-go terms. That is a different service, not evidence of FlyHermes cancellation behavior. Record which dashboard and invoice own your workspace before taking action.
Keep copies of useful delivered outputs and ask about export, retention and reauthorization before moving sensitive work. A hosted workspace does not automatically import local files or browser cookies. The hosted purchase checklist turns these questions into a bounded first-task evaluation. Terms checked November 2, 2026; no migration, cancellation, refund or recovery was performed for this comparison.
Compare the actual seller and current allowance, not a hosting label#
Nous Hermes Cloud and FlyHermes are separate services with different accounts and terms. The current official Nous hosting page, checked November 9, 2026, says paid plans include one Cloud agent up to the plan's size; extra or larger agents are billed by the day. Model and tool usage counts toward plan usage, with daily, weekly and monthly limits. This replaces the older assumption that every Cloud instance is necessarily an additional compute charge. Confirm your actual plan and deployment quote rather than applying an old FAQ to a new subscription.
The same official page says stopping an agent pauses compute but leaves a storage-only rate, while jobs requiring a running agent do not run. Stopping an instance, deleting it and canceling its subscription are different actions. These are Nous terms, not evidence about FlyHermes credits or retention.
FlyHermes's public offer, checked on the same date, advertises a hosted agent with included AI credits, Telegram, WhatsApp, Slack, Discord and web chat. Its monthly and annual allowances are different, as detailed below. It is independent of Nous Research. A plan called “Plus” in a community question does not establish which service the author bought; ask for the named seller and plan before recommending a migration.
Use the cost calculator worksheet to separate cash commitment, included usage, top-ups and infrastructure. Neither a higher allowance nor managed hosting establishes that your particular model/tool workflow is supported or faster.
Identify the failure before paying to move it#
A hosting purchase can remove server work. It does not, by itself, diagnose a stalled model response. Use this short triage before renting a larger VPS or moving a daily business workflow:
- Client cannot connect: compare the selected backend and profile, authentication and the live WebSocket. Use the remote Desktop guide, not a new provider subscription, to investigate an HTTP-ready but disconnected client.
- Ordinary chat works but a tool turn stalls: retain the model/provider, installed versions, timestamp and redacted tool result. Compare a harmless text-only turn with one read-only tool turn. Successful local execution is not proof that the model accepted the tool result. Follow the provider and rate-limit guide; do not call the symptom a credit shortage without the actual error.
- A scheduled report is missing: inspect the run and its intended destination separately. A completed file can coexist with failed channel delivery. The 24/7 acceptance guide keeps these checks separate.
- The host runs out of resources: inspect host RAM, disk and effective container limits, then size the actual workload. A remote-screen failure is not evidence that ordinary text chat needs the same capacity.
A November 8 community report describes local chat working while a tool turn hangs after producing output. That is an attributed, unresolved report, not a reproduced bug, proof of a provider outage or evidence that FlyHermes fixes it. The diagnostic checks above are a proposed evaluation method; we did not run the reporter's setup.
Choose a home hub, a VPS or a managed workspace#
If the goal is “the same agent from my laptop and desktop,” first choose one runtime to own the shared profile. Installing Hermes separately on both computers does not make their sessions, skills or files one shared workspace.
- Home server: a useful fit when the inputs live on your home network and you want to control the machine. Connect each client to that server and select the same server profile. You still own power, internet access, protected remote access, updates and recovery. Avoid this path if nobody can recover it while you are away.
- VPS: a useful fit when the runtime should not depend on your home's power or network. It remains self-hosted: you maintain the operating system, credentials, storage and services. A VPS does not automatically see a NAS or files on a laptop. Plan the input path before renting capacity; use the VPS deployment guide.
- FlyHermes: a useful fit when you want a managed workspace and supported browser/mobile or messaging access instead of operating a host. Confirm that the required accounts, files and tools are accessible through the service. Do not assume a private-network connection or the Desktop remote-backend method is included just because self-hosted Hermes supports it.
The official Desktop connection guide, checked October 19, 2026, scopes agents to their owning gateway and profile. Two profiles named default on different gateways are not the same workspace. Making a connection Primary does not switch the active Sessions workspace. Use the remote Desktop setup guide for the connection steps rather than copying secrets and skills onto every client.
Check shared state without assuming file sync#
On two clients, select the same named server connection and profile. Open a harmless existing session and compare its title and history. Then ask the agent to read a non-sensitive test file that you deliberately placed on the server. Check the returned contents. A file that exists only on the laptop is a separate input-access question; a successful remote connection is not a file-sync service.
Record the owning host and profile before adding schedules. Keep the home-server Docker state directory and the 24/7 recovery checklist with that record. This is a proposed acceptance check, not a claim that we tested your network or a managed account.
Monthly and yearly are different allowance decisions#
The FlyHermes public offer, checked November 9, 2026, lists $49 monthly with 490 AI credits each month, or $468 billed yearly ($39/month equivalent) with 4,680 credits for the whole year added at once. Top-ups start at $10. Check final checkout terms and taxes before paying.
For twelve months of monthly billing, multiply the monthly payment and credit allocation by twelve; those totals are not the annual plan's price or allowance. The yearly payment is $120 lower, but its advertised annual allocation is also 1,200 credits lower. The yearly balance is available upfront rather than through monthly allocations. This is arithmetic on advertised terms, not proof of equal usage or savings on your workload; do not assume rollover or convert credits into a fixed number of tasks.
Choose the billing period after estimating workload and cash commitment, not from the monthly-equivalent headline alone. The cost calculator worksheet separates power, infrastructure, model usage and operator time. It does not invent a maintenance-hours penalty to make managed hosting win.
Commercial disclosure: this site promotes FlyHermes. FlyHermes is an independent managed service, not Nous Research. Self-hosting remains the better fit when required local access or runtime control outweighs the maintenance you will own.
Include access removal in the hosting decision#
Onboarding is only half the operating model. Before handing a workflow to a teammate, decide who can remove that person, revoke a compromised credential and apply the resulting runtime change. Self-hosted Dashboard has real browser Chat and management controls, but the operator still owns this handoff.
The official Pairing page separates approved-user revocation from clearing pending requests. The Channels page says disabling a channel retains its token. Neither action is a substitute for revoking a compromised token with its issuer.
For a managed service, ask who performs each action, what support is included, what data can be exported and which external accounts remain yours. Verify those answers against your actual plan; do not assume managed hosting supplies every team-access or offboarding feature. The dashboard runbook is the self-hosted comparison baseline.
Ask who owns the service conflict and the interrupted task#
Two practical questions expose the difference between a convenient interface and managed operation.
Who fixes a refused service start? Current self-hosted Dashboard documentation describes exit 78 when an existing backend owns the host but cannot satisfy a new launch's address settings. Someone must identify that owner, reconcile supervisors and decide whether to reuse it or deliberately replace it. The dashboard runbook now includes that handoff; buying another port or restarting indefinitely does not supply an operator.
Who reconciles work interrupted at the client? The same documentation distinguishes a reconnecting WebSocket from closing the browser Chat tab, which reaps its PTY. Resuming history does not prove a command continued. Define who checks the actual file or external record before permitting a retry. A managed service may own the host, but the buyer still needs task-completion and approval rules.
Before choosing FlyHermes, write down one representative job, its required accounts/files, supported channel, expected usage and escalation path. Confirm export and recovery terms rather than inferring them from a dashboard demo. Its public offer includes a hosted runtime, supported model access and a plan-specific usage allowance; that is not an unlimited-inference or every-plan SLA guarantee. These buying checks follow the current official Dashboard reference and the managed offer, not a claimed outage or benchmark.
Choose the operating model, not the window#
Self-hosted Web Dashboard: you operate the installation#
The Hermes Agent Dashboard manages configuration, profiles, keys, sessions, logs, skills, tools and scheduled work. Current official documentation also describes a Chat tab that embeds the real terminal UI over a PTY/WebSocket. That is a self-hosted interface, not the FlyHermes hosted chat product. The current Dashboard reference requires Linux, macOS or WSL2 for this POSIX PTY pane; native Windows can use the management pages but shows a WSL2 notice for browser Chat. Check your installed build before choosing that surface.
A dashboard can make administration easier without taking responsibility for the machine. You still decide who can reach it, which credentials it holds, how it restarts and how state is restored. Start with the private Web UI setup guide if those are responsibilities you want.
Hermes Desktop: a client connected to a particular backend#
Desktop is the native workspace. Its selected gateway and profile determine where the session's files, memory and tools live. Connecting Desktop to a VPS does not copy your laptop's files or browser logins there. Nor does making a connection Primary necessarily switch the current Sessions workspace.
Use the remote Desktop connection guide when you want the native client but intend to maintain the remote runtime. A connection test and a harmless session on the intended profile are more useful than matching the version labels on two windows.
FlyHermes: a managed Hermes workspace#
FlyHermes advertises a hosted runtime, browser chat, supported model API access, tooling infrastructure, core channels and a plan-specific usage allowance. Its value is removing infrastructure assembly and day-to-day runtime maintenance, not making the software open-source in a different way.
Before subscribing, confirm the current supported channels, model access, allowance and account terms on the FlyHermes offer. Ask separately about private-network connectivity, data export, required integrations and team access. Do not infer support for a particular Desktop connection method, Bot Screen, SSO or custom deployment from the word “hosted.”
Match the choice to one real job#
A morning research brief: managed hosting is a good fit when the inputs are accessible to the hosted agent and the outcome is a report in your preferred supported channel. You still define the sources, cutoff time and what counts as a useful result. Self-hosting can be a good fit when the brief needs a private database or files that must remain inside your network.
An inbox assistant: start with draft-only work, whichever hosting model you choose. Verify the connected mailbox, authorized scope and approval step before permitting sends. The October 4 Sharbel workflow walkthrough explicitly recommends draft mode rather than automatic sending. That is a useful workflow constraint, not evidence of guaranteed sales or accuracy.
Work from your phone: a phone is the client, not proof of an always-on backend. If Hermes runs on a sleeping laptop, installing another mobile interface does not keep the runtime awake. The phone access guide separates messaging, private dashboard access and managed web chat.
Browser work requiring a login: check where the browser actually runs and how you will complete login or approval challenges. A hosted agent cannot automatically inherit the cookies from your personal computer. Require a small read-only demonstration against the intended account before giving it a consequential task.
Do not buy a larger server for the wrong failure#
“Not enough memory” can refer to host RAM, a container limit, model context or an optional remote screen. Those are different problems.
A September 30 community support thread initially described trouble connecting Desktop to a cloud instance. The discussion identified Bot Screen, not ordinary dashboard administration. The user later reported that a larger instance allowed the screen to connect. This is one support case, not a universal sizing benchmark or a FlyHermes plan recommendation.
The official Bot Screen documentation explains why: a Linux desktop and headed browser run on the owning host or supported sandbox, with a free-memory admission check. That workload is different from opening an admin page. Identify the exact pane and failing operation before buying capacity or changing a model.
For a self-hosted buying decision, record:
- The required job: text chat, a scheduled report, browser interaction or remote screen takeover.
- The machine or container that executes it, not just the device showing the interface.
- Available memory and the effective container limit during the failed attempt.
- The exact error and timestamp, with secrets and private paths removed.
- Whether the basic chat still works when the optional screen is closed.
The VPS setup guide covers the operating checks. If maintaining that inventory is not work you want to own, compare managed hosting—but still confirm that the managed offer supports the exact workflow.
What each side still has to own#
Runtime, updates and recovery#
With self-hosting, you own the server or dedicated computer, process supervision, storage, updates and rollback. Docker changes packaging; it does not supply an operator. A running container can contain a failed gateway, and a recently updated Desktop can be ahead of its backend. Use the Docker troubleshooting guide to identify the failed layer before rebuilding anything.
With managed hosting, the service operates the hosted runtime. You still need an escalation route and a way to determine whether a missed result was a platform outage, a rejected model call, a tool permission failure or an incomplete task. Ask what support and recovery the actual plan includes; do not assume a guaranteed response time or SLA.
Credentials, permissions and approval#
Self-hosters own dashboard authentication, secrets, gateway allowlists and access to connected services. Keep administrative access distinct from the channel where a teammate asks for a report. The server security checklist covers private access and public authentication boundaries.
A managed service can handle infrastructure and supported model access, but it cannot decide which business records each user should see. You own account authorization, least-privilege scopes and approval for actions such as sending email, publishing content or making payments. Ask about role separation before using one shared assistant for unrelated clients.
Usage, allowances and the bill#
Self-hosted software, infrastructure and model inference are separate cost lines. Include paid tools, storage, backup and the time spent diagnosing failures. Do not estimate the whole system from the price of the smallest VPS.
FlyHermes includes model access and a plan-specific usage allowance in its advertised offer. Review the current allowance, supported models and exhaustion behavior; neither “included” nor “managed” means unmetered. For variable inference spend, use the provider-cost measurement guide. Compare the same representative job and accepted result on both paths rather than assuming one is always cheaper.
State and exit options#
For self-hosting, test restoration of a protected backup before trusting it. An export is not a tested recovery, and a copied profile name is not proof that its credentials, files and schedules work on another host.
For managed hosting, confirm what can be exported, what must be reauthorized and how cancellation affects access or retention. Do not assume automatic migration of private files, browser cookies or running jobs. Keep a rollback option until one representative workflow has passed on the new runtime.
There is also a separate implementation-service purchase: paying a team to scope, integrate and launch one business workflow. The Agentra review examines that model and its contract-defined handoff. Do not compare a managed Hermes hosting fee with a deployment-pod quote as if both include workflow discovery, finance approvals, user training and acceptance testing; first write down which responsibilities you need another party to own.
Self-hosted tools can coexist with a vendor-operated agent control plane. The Claude Managed Agents review explains why local tool execution is not equivalent to local inference, zero retention or an offline agent. Draw the data-flow boundary for inputs, tool results and session history before accepting a hosting label as a privacy guarantee.
The OpenAI Agents API review provides another concrete boundary: a self-hosted sandbox does not make the managed API eligible for Zero Data Retention. Separate the location of tool execution from the location of session state and orchestration. Hosting only one layer yourself does not establish an entirely local data path.
A free coordination platform can still leave the agent running on your own machine. The Agency Label Academy review separates a free workspace from the runtime, model tooling and human approval it depends on. Map each component before outsourcing maintenance: who runs the listener, who holds credentials, who approves previews and who restores a failed site? Buying managed agent hosting does not automatically manage a separate agency platform.
Run a bounded acceptance test before moving real work#
Use one small task that reads public information and produces a local draft or private report. Do not test recovery with a payment, public post or customer email.
- Name the backend and profile that own the task. Check its actual input access.
- Run it once and inspect the output, not only the assistant's completion message.
- Close the client. For a scheduled workflow, verify a later result while that client remains closed.
- Check the intended destination and timestamp. A job row and a delivered report are separate facts.
- If a run stops unexpectedly, inspect its output and history before retrying. An uncertain completion can hide a finished external action.
- Record who will diagnose the next failure and what evidence they need.
The 24/7 setup guide expands the reliability test. This acceptance test is a proposed evaluation method, not a benchmark we claim to have run on your deployment.
Compare a working week, not a demo#
Keep a short record for each candidate setup. No invented hourly rate or monthly saving is needed:
Workflow and owner:
Backend / profile / client:
Required files, accounts and channel:
Accepted results / attempted runs:
Model and tool charges observed:
Infrastructure or subscription charge:
Operator time and incidents:
Approval and recovery steps:
Export or rollback test:
Include failed attempts and time spent on authentication, updates, browser recovery and missed delivery. A low invoice can be a poor deal if you cannot maintain the service; a managed subscription can be a poor fit if the required private integration is unavailable. Compare actual coverage before price.
Move one workflow without running it twice#
Inventory the old task, input locations, permissions, schedules and delivery destinations. Back up self-hosted state and confirm the new service's import or manual recreation path. Reauthorize only the accounts the new workflow needs; never paste credentials into a migration document.
Test on the new host with external side effects disabled. Check that the expected profile, files and channel are in use. Pause the old schedule before enabling the replacement, then verify one new run and its destination. Keep the old setup available for a time-boxed rollback without leaving both schedules active. Retire old credentials and infrastructure only after the new path is accepted.
The decision#
Choose self-hosted Hermes when control is the requirement and someone can own the operating work. Choose FlyHermes when the requirement is a managed assistant you can reach from browser or phone, with supported model access and channels, and the current offer covers your workflow.
Start with FlyHermes pricing and plan checks, or review the managed workspace. If you choose self-hosting, use the linked dashboard, security and VPS guides rather than treating a successful install as the end of setup.
Product boundaries were checked against the official Web Dashboard guide, Desktop connection guide and FlyHermes's public offer on November 9, 2026. The official Nous hosting offer was also rechecked that day; these are rolling references, not a pinned release test. Community and creator examples explain the questions buyers ask; they are not guarantees of compatibility, performance or financial results.