Self-hosted agents on your AI plan: OtoDock, Zylos and Sentō

OtoDock, Zylos and Sentō document ways to run agents on infrastructure you control using an existing AI subscription. Choose by the operating model you want: a multi-user platform, a persistent cross-channel assistant, or a Claude Code agent reached through messaging. Self-hosting does not make a cloud model private, remove its limits or authorize sharing one personal account across a team.

Checked 6 November 2026. These are documentation-based comparisons, not tested deployments or compliance certifications. Browse the listings for OtoDock, Zylos and Sentō, then compare the broader Claude and ChatGPT directories.

Which self-hosted agent matches your operating model?

App	Documented starting point	Decision to settle before installation
OtoDock	Claude Code and Codex connections, with separate API and local-model options	Which work runs on personal connections versus platform engines?
Zylos	Persistent assistant with Claude Code or Codex runtimes	Who operates the host, messaging channels and recovery process?
Sentō	Claude Code agent reached through Discord, Telegram, Slack or iMessage	Which channels and host capabilities should the bot receive?

Start with the smallest setup that fits the job. A personal daily research assistant and a shared company workflow have different identity and permission needs, even if both can run the same model.

The existing business-agent comparison covers task fit across hosted and self-hosted products. This page concentrates on the operator's responsibilities: host resources, credential boundaries, unattended permissions and recovery.

What should you inspect in OtoDock before enabling shared work?

OtoDock's connection guide distinguishes personal engines under User Settings from platform engines under Setup. It documents Claude subscription sign-in and an OpenAI device-code flow for Codex. If OpenAI refuses device login, the guide directs you to the corresponding ChatGPT security setting.

An administrator can contribute a personal connection to platform work. The guide says that contribution is enabled by default. Inspect it before scheduling tasks; the last person chatting with an agent is not necessarily the person whose connection funds agent-owned work.

That is an application setting, not permission from a provider to pool consumer subscriptions. Do not roll out a shared account because a dashboard offers a checkbox. Establish the identity and funding arrangement first.

Read the permissions guide separately. Interactive modes and unattended tasks are different control surfaces. The guide says scheduled tasks do not pause for approvals. Begin with work that only reads public material and saves a local draft, with no purchasing, external messages or production writes. This is a suggested pilot, not a report of running it.

What does Zylos require from the person operating it?

Zylos's homepage describes Claude Code and Codex as interchangeable runtimes, with memory, skills and channels preserved when switching. Its installation guide is narrower: the prerequisites and authentication examples focus on Claude. Confirm the current Codex setup route before assuming every install command applies unchanged.

The installation page lists Linux or macOS, at least two cores, 4 GB of RAM and 20 GB of disk, plus Node.js 20.20 or later. These are maker requirements, not measured capacity for your workload. Leave room for files, browser work and logs rather than buying a server exactly at the minimum.

The guide provides a doctor command for dependency, authentication and service checks. Make that part of your recovery checklist, alongside a backup of persistent agent data and a way to stop the service outside the agent itself. Do not paste tokens into a public support issue when troubleshooting.

An always-on scheduler adds operating responsibility. Decide what happens when the host reboots, a login expires or a task repeats after interruption. A successful first chat does not demonstrate recovery from any of those failures.

What needs checking before you deploy Sentō?

Sentō's setup guide documents Claude subscription authentication during installation and a separate token-preparation route for Docker. Its README describes a Claude Code-based agent controlled through messaging services.

Do not translate the README's unlimited-usage language into a provider promise. Running software continuously does not create more subscription allowance. Check eligibility in the provider account, especially for organization-managed seats; the maker's broad plan list is not a substitute for your administrator's controls.

The setup guide also identifies a host-specific boundary: iMessage uses the Messages app's local database on macOS rather than an official Apple API. If that channel matters, a generic Linux VPS is not an equivalent deployment. Review the host and account access required before choosing it.

Give the bot a narrow channel scope and avoid production credentials for the first assignment. A useful acceptance test is generating a draft from public pages, saving it, restarting the service and checking whether it can retrieve the saved result. Record what you actually tested rather than assuming every advertised feature works on your host.

What do provider rules and self-hosting costs still require?

Anthropic's legal guidance permits an end user to sign into the unmodified Claude Code binary, including in a hosted platform under its stated conditions. It separately prohibits third-party developers from collecting or intermediating Claude.ai credentials or routing requests through consumer credentials on users' behalf. Each end user must authenticate under the documented hosted-binary arrangement.

These distinctions matter for tools that expose subscription-token fields or shared pools. We have not audited these products' credential handling. Verify the installed runtime and sign-in route instead of treating a maker's support claim as a compliance finding.

Budget for the host, backups, storage and any external services as well as model access. Keeping the orchestration layer on your machine does not keep prompts there when it calls a cloud provider. Local-model inference is a separate option with different hardware and quality trade-offs.

Before leaving an agent unattended, confirm five things: the paying account, permitted tools, recoverable data, a task spending boundary where available, and an independent stop mechanism. If all you need is access to an existing coding session while away, phone control of Claude Code and Codex is a narrower problem than operating a persistent agent server.