VAOS vs Running OpenClaw Yourself

OpenClaw is open source under the Apache 2.0 license. You can clone it, run it on your own infrastructure, and modify every line. Nothing about VAOS changes that. This post is a straightforward comparison of what each option involves, where each one wins, and when you should pick one over the other.

What self-hosting OpenClaw looks like

The process is not complicated, but it is not trivial either.

1. Clone the repository. Install Elixir, Erlang/OTP, and project dependencies via Mix.
2. Configure a provider. Set API keys for your LLM of choice (OpenAI, Anthropic, Zhipu, etc.).
3. Set up a Telegram bot token if you want the Telegram channel.
4. Deploy to a server. Fly.io, Railway, bare metal, a Raspberry Pi -- your choice.
5. Configure environment variables, run migrations, start the application.

Estimated setup time: roughly 60 minutes if you are familiar with Elixir. Longer if you are not. Elixir's toolchain is well-documented, but OTP concepts like supervision trees and GenServers have a learning curve that is measured in weeks, not hours.

Once running, you manage: server uptime, application updates (pulling new releases, running migrations), scaling under load, monitoring for crashes, and backups. Standard ops work.

There is one architectural detail worth calling out: agent memory is ephemeral. When the BEAM process restarts -- planned deployment, crash, server reboot -- the agent starts with a clean slate. It forgets every correction you made, every behavioral rule it learned, every piece of context from prior conversations. You can build persistence yourself, but the open-source release does not include it.

What VAOS managed hosting looks like

1. Sign up. Pick a plan: $29/month (Starter), $49/month (Pro), or custom pricing for enterprise.
2. Configure your agent through the dashboard: select a model, connect a Telegram channel, set the system prompt.
3. Deploy. One click. The agent is live on Fly.io in approximately 60 seconds.

After that, you manage your agent's behavior. We manage everything else -- infrastructure, uptime, deployments, monitoring.

The meaningful difference is not convenience (though that matters). It is what VAOS adds on top of the open-source core:

1. Persistent memory layer. Corrections and behavioral rules survive restarts. You tell the agent "do not respond to pricing questions with estimates" once. It sticks. Restart, redeploy, scale up -- the correction persists. This is the single largest operational difference between self-hosted and managed.
2. Dashboard for agent configuration. Model selection, system prompt editing, channel setup, behavioral corrections -- all through a web UI instead of config files and redeployments.
3. One-click Fly.io deployment. No Dockerfiles, no fly.toml tuning, no release configuration.
4. Opik tracing integration. Full observability into agent behavior -- every LLM call, tool invocation, and routing decision is traced and inspectable.
5. Token wallet billing. A wallet system tracks token usage per agent. You load credits, the system deducts per inference call. No surprise bills from runaway loops.
6. Checkpoint/restore for agent state. Snapshot and roll back agent state if a configuration change goes wrong.

The comparison

The cost delta is real. Self-hosting on a $5/month Fly.io instance is $24-44/month cheaper than VAOS before token costs. If you are running a single agent for a personal project and you know Elixir, self-hosting is the obvious economic choice.

When VAOS is the right choice

• You want persistent memory without building a persistence layer yourself. This is non-trivial to implement correctly -- it involves state serialization, storage, rehydration on restart, and conflict resolution when corrections overlap.
• You do not want to manage Elixir/OTP infrastructure. If your team does not have Elixir experience, the operational burden of keeping a BEAM application healthy in production is a real cost that does not show up in the hosting bill.
• You need quick iteration on agent behavior. Correcting an agent through the dashboard takes seconds. Correcting it through code changes requires a commit, a build, and a deployment.
• You want monitoring without setting up an observability stack. Opik tracing is included. No Prometheus exporters, no Grafana dashboards, no alerting rules to configure.

When self-hosting is better

This section is not a hedge. These are cases where self-hosting is genuinely the better option.

• You need full code-level customization. If you want to modify signal classification logic, change routing behavior, add custom tools, or alter how the agent reasons, you need the source code and the ability to deploy your own builds. VAOS does not expose these internals.
• You are building the agent into a larger Elixir application. If OpenClaw is one component in a bigger OTP system -- embedded in a Phoenix app, supervised alongside other GenServers -- self-hosting lets you run it as a dependency rather than a separate service.
• You want to modify the core. Signal classification, tool routing, prompt construction -- these are all in the open-source code. If your use case requires changes to these systems, self-hosting is the only path.
• You are cost-sensitive and comfortable with Elixir ops. $5/month versus $29/month is a meaningful difference when you are running multiple agents or operating on a tight budget. If you can handle the ops, the savings are straightforward.
• You need features VAOS does not support yet. Custom tool registration, multi-node agent coordination, and non-Telegram channels (Discord, Slack) are available in the open-source release but not yet exposed through the VAOS dashboard.

What VAOS does not do yet

Transparency on current limitations:

• Custom tool registration. You cannot add your own tools through the dashboard. If you need tools beyond the built-in set, you need to self-host.
• Multi-node agent coordination. Running agents across multiple nodes for redundancy or geographic distribution is not supported in the managed platform.
• Non-Telegram channels via dashboard. Discord and Slack integrations exist in the CLI tooling but are not configurable through the web dashboard yet.
• Self-hosted VAOS. The persistence layer, dashboard, and tracing integration are managed-only. There is no "install VAOS on your own server" option. You get OpenClaw (open source) or VAOS (managed). Not both in one package.

The short version

OpenClaw is the engine. VAOS is the engine plus persistent memory, a dashboard, monitoring, and managed infrastructure. If you want control, self-host. If you want your agent to remember what you told it yesterday without building that yourself, use VAOS.

Both options use the same core. Choosing one does not lock you out of the other.

OpenClaw is available at github.com/openclaw under the Apache 2.0 license. VAOS plans and pricing are at vaos.sh.