AI Agents
A long-lived AI assistant of your own, reachable from its web UI and from your chat apps, run by SailOps on a hardened runtime with its memory on a persistent volume.
An AI agent is a managed service, like a database: you choose an engine and a size, SailOps runs one instance of it, publishes its web UI over HTTPS, keeps its state on a volume and bills it by the hour. Agents live under AI › Agents in the project console.
Four engines are offered. They are upstream open-source projects; SailOps packages and operates them. The Nanobot image patches dependency bounds to use corrected Dulwich and pypdf releases.
| OpenClaw | Hermes Agent | |
|---|---|---|
| What it is | A personal assistant gateway with a web Control UI, tools and skills | A self-improving agent with a web dashboard, persistent memory and skills |
| Web UI | Control UI, protected by a gateway token | Dashboard, protected by a login (admin and a password) |
| Channels at launch | Telegram | Telegram, Discord, Slack |
| State on the volume | /home/node/.openclaw | /opt/data |
| Minimum flavor | 1 vCPU, 2 GiB | 0.5 vCPU, 2 GiB |
| Additional engine | Nanobot | ZeroClaw |
|---|---|---|
| Version | 0.3.5 | 0.8.6 |
| Web UI | WebUI, protected by an access token | Dashboard, paired using a browser code |
| Channels | Telegram, Discord, Slack | Telegram, Discord, Slack |
| State | /home/agent/.nanobot | /home/agent/.zeroclaw |
| Minimum resources | 0.5 vCPU, 1 GiB | 0.5 vCPU, 1 GiB |
Nanobot and ZeroClaw use the same compute catalogue and billing as the other engines. The initial minimum is a supported deployment floor, not a measured promise about peak consumption. Their channel libraries and dashboard are packaged in the managed image. Tool execution uses the pod's gVisor isolation, without a Docker socket or nested container runtime.
The engine is chosen when the agent is created and cannot be changed afterwards.
An agent acts for whoever holds its credential: that person drives its tools, its channels and everything it remembers. Do not share one agent between people who should not see each other's conversations. Create one agent per person or per team instead.
Create an agent
- Open AI › Agents and click New agent.
- Name it (lowercase letters, digits and hyphens, 40 characters at most) and pick OpenClaw, Hermes Agent, Nanobot or ZeroClaw.
- Choose the model provider: SailOps Models (the default) or another OpenAI-compatible provider. See Model provider.
- Paste the channel tokens you want, if any. See Channels.
- Choose a compute class and a flavor. Only the flavors that meet the engine's minimum are offered, each with its price per month. Choose the size of the state volume (5 GiB by default, 1 to 500 GiB) and its class.
- Optionally restrict the web UI to a list of client IPs or CIDR ranges.
- Click Create agent. The console opens the agent's Credentials page: reveal the credential and store it now.
The agent starts as soon as its image is pulled. It is published on an unguessable hostname of the shared gateway, shown as Open Control UI or Open dashboard on its page.
Managed agents are not available during the trial: an organization in trial gets a refusal, and the console shows Paid organization required. There is no free tier.
Approve your OpenClaw browser
After entering the gateway token, OpenClaw displays a browser approval request. Return to Credentials in SailOps, click Refresh browser requests, and compare the request ID with the one shown in OpenClaw. Click Approve browser and confirm only your own request. Return to OpenClaw and reconnect. You do not need a server terminal. This step is not needed for Hermes.
Credentials
SailOps generates the credential that protects the web UI when the agent is created:
- OpenClaw: a gateway token. The Control UI asks for it once per browser.
- Nanobot and ZeroClaw: an access token to paste into the web interface (Nanobot calls this its password). No OpenClaw browser approval is needed.
- Hermes Agent: the username
adminand a password, for the dashboard login.
The credential is shown once. Reveal credentials on the Credentials page returns it the first time; every later read returns the URL and the username, but not the secret. Store it in your password manager.
Reset generates a new one, shows it once and restarts the agent. The old one stops working, and every browser signed in with it has to sign in again: OpenClaw also forgets every browser paired with its Control UI when it first starts with the new token. A reset is the only way to obtain the secret again.
Revealing and resetting require the ai-agents:update scope (DEVELOPER and above), and both are recorded in the audit log.
Model provider
An agent calls a language model through an OpenAI-compatible endpoint: a base URL, a model and an API key.
SailOps Models (default)
With SailOps Models, the agent calls SailOps Models at https://app.sailops.io/api/ai/v1. SailOps creates a project API key labelled models-agent.<agent name>, holding the ai-models:invoke scope only, and writes it into the agent: there is no key to paste. Leave the model empty to use the default one (qwen3.8-27b), or name any model listed on the Models page.
Use SailOps Models on the agent's overview does the same for an existing agent: it creates the key, or rotates it if it exists, points the base URL at SailOps Models and restarts the agent. It replaces the API key the agent had, which cannot be read back. The previous Models key is revoked only once the new one is in place, so a binding that fails leaves the agent working. Tokens are billed to the project by SailOps Models, within its monthly budget. Deleting the agent revokes its key.
Through the API, binding mints a key holding ai-models:invoke, so the API key that asks for it needs ai-models:invoke as well as ai-agents:update. Bosco never runs it: ask it and it points you to the console.
Another OpenAI-compatible provider
Choose Another OpenAI-compatible provider and enter its base URL (for example https://api.openai.com/v1), the model identifier it serves, and its API key. Any endpoint that speaks the OpenAI chat completions API works. The base URL and the model can be changed later from Settings; the key from Model key and channels.
Moving the base URL to another host, scheme or port clears the stored API key, and revokes the agent's SailOps Models key if it had one: a key is never sent to a host it was not given for. Set the new provider's key from Model key and channels afterwards. A different path on the same host keeps the key.
Channels
Besides its web UI, an agent can answer on chat apps. Each channel is enabled by a token:
| Channel | Token | OpenClaw | Hermes Agent |
|---|---|---|---|
| Telegram | Bot token from @BotFather | ✓ | ✓ |
| Discord | Bot token from the Discord developer portal, with the message content intent | — | ✓ |
| Slack | Bot token (xoxb-…) and app token (xapp-…, Socket Mode) | — | ✓ |
OpenClaw installs Discord and Slack as plugins at runtime, which a managed agent does not do: they are not offered for OpenClaw yet.
The first time someone writes to the bot, the agent answers with a pairing code; approve it from the agent's web UI. Until then, strangers cannot talk to your agent.
Tokens are write-only
The model API key and the channel tokens are write-only. They are stored in a Kubernetes Secret of the agent, read by the agent's pod, and never returned: not by the console, not by the API, not by Bosco. The agent's page only tells you which ones are set. On Model key and channels, type a new value to replace one or Remove it to clear it; saving restarts the agent. Secrets never go through the MCP tools either: set them from the console or the API.
If that Secret is ever lost, SailOps recreates it with a new credential but cannot recreate your keys and tokens, which existed nowhere else. The agent then shows Warning until you set them again, or run Use SailOps Models.
What lives on the volume
Everything the agent accumulates is on its state volume: conversations and sessions, memories, installed skills, channel pairings and its workspace. The volume survives stops, restarts and updates, and is deleted with the agent.
The engine's main configuration file is managed by SailOps and written again at every start from the agent's settings: changes made to that file from the agent's own settings screen are replaced at the next restart. Change the model, the provider and the channels from the console instead.
Start, stop, restart, delete
- Stop stops the agent's compute. Its web UI and channels stop answering, its volume is kept, and only the volume is billed.
- Start brings it back with everything it remembered.
- Restart restarts its pod, for example after changing something from its own UI.
- Delete (project OWNER) removes the agent, its volume, its credential and its secrets, and revokes its SailOps Models key. It cannot be undone.
A change is accepted while a previous one is still rolling out: an agent that never becomes ready, because of a revoked key or a refused channel token, can always be fixed from the console.
Updates
SailOps watches both projects and moves to their new releases after review. When a newer version is available, the agent's page says so. Settings › Update to … (or saving any change from Settings) moves the agent to it: the agent restarts on the same volume with the new image. There is no way back to the previous version once the new one has started on the volume, so export what matters from the agent's own UI first.
If the agent does not come back after an update, open its Logs. OpenClaw exits with code 78 when its configuration or its state needs a migration it will not apply on its own; the agent then shows Error. Contact support with the agent's name: the volume is kept.
Skills and ClawHub
OpenClaw installs skills from ClawHub, and Hermes Agent from its own skill registry. SailOps does not audit these skills. A skill runs with the agent's permissions, inside its pod, with the agent's keys and tokens available to it. Install only skills you have read, from authors you trust.
Pricing
An agent is billed like the other managed services: its compute by the hour while it runs, at its compute class price, plus its state volume. Model tokens are billed separately, by SailOps Models or by your provider.
List prices per month of compute, excluding VAT. The console shows your organization's prices on each flavor.
| Class | Flavor | vCPU / memory | Per month | Engines |
|---|---|---|---|---|
| eco | medium | 1 / 2 GiB | 31.25 € | OpenClaw, Hermes Agent |
| eco | large | 2 / 4 GiB | 62.50 € | OpenClaw, Hermes Agent |
| eco | xlarge | 3 / 6 GiB | 93.75 € | OpenClaw, Hermes Agent |
| standard | small | 1 / 4 GiB | 83.50 € | OpenClaw, Hermes Agent |
| standard | medium | 2 / 8 GiB | 167.00 € | OpenClaw, Hermes Agent |
| production | small | 1 / 4 GiB | 112.50 € | OpenClaw, Hermes Agent |
| production | medium | 2 / 8 GiB | 225.00 € | OpenClaw, Hermes Agent |
Larger flavors are listed in the console. The state volume costs about 0.12 € per GiB and per month in the standard class (0.62 € for the default 5 GiB) and 0.17 € in the performance class, and stays billed while the agent is stopped.
Security
- The agent runs under gVisor, as a non-root user, with every Linux capability dropped, no privilege escalation and no Kubernetes service-account token.
- Its pod accepts traffic only from the shared gateway's proxies, on the web UI's port. Other workloads of the project cannot reach it directly, and Hermes Agent's own API server is not exposed.
- The web UI is served over HTTPS on an unguessable hostname, always behind the agent's credential, and optionally behind a client IP allowlist. SailOps refuses to deploy an agent without a credential.
- Secrets live only in the agent's Kubernetes Secret and reach the pod by reference.
- The agent keeps the project's default outbound access (the Internet and the project's own services), which its tools and channels need. Its shell tools run inside its own container.
API and MCP
The public API serves agents at https://app.sailops.io/api/public/api/v1/ai/agents with a project API key: list, create, get, update, delete, start, stop, restart, credentials and their reset, secrets, the SailOps Models binding, flavors and engines. Reads need the ai-agents:list scope, every other operation ai-agents:update.
The SailOps MCP server exposes the services_ai_agent_* tools (list, create, get, update, delete, start, stop, restart, credentials reset, Models binding, flavors, engines, logs and metrics) to Bosco and to your own agents. They never carry the model API key or the channel tokens. services_ai_agent_credentials_reset is the one tool that returns a secret: the new web UI credential, once, in its result. Bosco runs neither the reset nor the Models binding.
Nanobot and ZeroClaw channels
Set Telegram, Discord and Slack tokens through SailOps Channels, then manage approved senders from the engine's authenticated web interface. Both Slack tokens are required to enable Socket Mode. Nanobot starts with allowlist/pairing policies for DMs and groups; ZeroClaw keeps its upstream peer approval defaults. Adding a bot token does not authorize every sender. Test a direct message and approve only your own sender before inviting a team.
SailOps owns the model endpoint and channel credentials configured in the console. Saving settings or restarting reapplies those managed fields. Conversations, memory and workspaces remain on the state volume.
For ZeroClaw, open Credentials → Generate browser code, then paste the code into the dashboard pairing screen. The revealed access token is for API access. Paired browsers must pair again after a rollout; resetting credentials also revokes their old access.