Wyrmling
A local AI agent that runs on my own hardware, with jailed tools, scheduled routines, and automatic failover from a GPU machine to a CPU machine.
Wyrmling is an AI agent that runs entirely on my own machines, using open-weight Qwen3 models rather than a cloud service. I talk to it through a web chat or Discord, and it also runs routines on a schedule.
- Web chat connects to Wyrmling daemon
- Discord bridge connects to Wyrmling daemon
- Scheduled routines (checked every 30 s) connects to Wyrmling daemon
- Wyrmling daemon connects to Tool jail (sandboxed, allowlisted network)
- Wyrmling daemon connects to Chat history and routines (local files)
- Wyrmling daemon connects to Model front door (failover proxy)
- Model front door (failover proxy) connects to GPU machine (preferred) (first choice)
- Model front door (failover proxy) connects to Always-on CPU machine (backup) (after a failure)
How it fits together
The daemon keeps every conversation as a file on disk and checks for due routines every 30 seconds. When the model asks to use a tool, the tool runs inside a sandbox that can only see its own workspace and can only reach an allowlisted set of network addresses.
Every request to a model goes through one front door, a proxy that prefers the GPU in my main PC. If that machine fails to answer, the proxy marks it down for 30 seconds and sends requests to a slower CPU copy of the models on an always-on Linux machine. Nothing that calls the model needs to know which machine answered. A request that the GPU machine has already started is not retried elsewhere, so a failure partway through shows up as an error rather than as a silent second answer.
An honest number
On 17 October 2026 I timed one long prompt on both paths. Launched by hand, the GPU machine generated 15.0 tokens per second against 7.8 on the CPU machine. Started the way it starts at boot, the GPU machine managed only 3.0, because the boot task runs at low priority and pages memory. So the failover works, but the speed advantage is not yet reliable, and I do not claim it is.
How it was built
The code is private and this site never connects to it. On 9 October 2026 its full test suite passed 798 of 798 tests, including 90 for the tool sandbox.