Agent infrastructure · 2026
Unattended agent infrastructure
In production
- Role
- Operator and engineer
- Stack
- OpenClaw, Hermes, MCP, LiteLLM
- Status
- In production
Problem
Chat UIs that call one model and forget the tools are not an operations stack. Clients who need agents that start on a schedule, finish a job, and say when they failed do not get that from a vendor playground.
Constraint
This stack runs on the same host as this site. Internal URLs, keys, and third-party channels stay off the public page.
Architecture
- Cron
- OpenClaw
- MCP tools
- LiteLLM
- Model vendors
- Alert on failure
Decisions
OpenClaw as the runtime
Agents, skills, and tool calls in one place that can run unattended. Not a notebook. Not a chat tab that dies when the laptop sleeps.
MCP for tools
Each capability is a server with a contract. The agent does not scrape ad-hoc APIs from a prompt.
LiteLLM in front of vendors
One proxy, failover when a model or a key dies. The job does not inherit a single vendor's outage.
Cron for the unattended path
Scheduled work is a first-class path, not a human clicking Run. Failures have to surface; they are not a quiet log line.
I will not publish a dashboard screenshot of private jobs. The proof is that this site and the agent stack share a machine, and that the architecture is specific enough to copy into a scope.
A typical starting point is an existing model vendor, a list of tools that are only connected in theory, and a job that has to run when nobody is watching.
Outcome
Running on this host. The same shape is what I install for a client who already has a model vendor and a set of half-connected tools.
What was handed over
- Architecture notes and what is in or out of scope
- How to start, stop, and fail over the proxy
- Skill list and which MCP servers they call
- Where secrets live — keys are not in the repo and not on this page