AI operations for network & infrastructure
WeftRoute runs a closed loop over your infrastructure. It drafts the plan, executes it under your approval, verifies the result against the state you approved — and when reality disagrees, it investigates and tells you what broke, with the evidence.
Why a loop
That's the moment the real risk starts. A workflow can return success while an uplink is down, a neighbour never came back, or a change landed on four devices out of thirty. WeftRoute treats the interval between "the engine said SUCCEEDED" and "the infrastructure is actually in the approved state" as work in its own right — and puts an agent on it.
Three agents, one loop
Describe what you want — an upgrade across the branch fleet, a compliance sweep, a config rolled back. It reads your inventory, your history and your own documentation, then drafts the plan.
Its only write path is a proposal.
You approve it, and it is re-validated at that moment against live state — not against the state it was drafted from.
After execution it probes real devices, compares them to the state you approved, and reconciles the difference. Only idempotent actions run on their own; anything else escalates to you with the device and the exact failing check named.
The engine reported SUCCEEDED.
The Operator disagreed, named the device, reconciled, and escalated.
On an alert it scopes the incident, gathers evidence, correlates against recent changes, forms hypotheses and scores them. An uncorroborated hypothesis can never be the root cause, and "inconclusive" is a real answer it is allowed to give.
Read-only by construction.
The tier ceiling is a constant in the code, not a setting — there is no path by which it writes.
Governance
Enterprise AI fails on trust, not capability. WeftRoute is built so the interesting question — "what happens when the model is wrong?" — has a boring answer.
Underneath
Everything an agent can do, you can do yourself on a canvas — palette nodes wired into flows, parsers saved as data, no scripting at any step. The AI drafts into the same engine your team edits by hand, which is why its output is reviewable rather than opaque.
Cisco IOS / NX-OS
Juniper Junos
Arista EOS
Nokia SR Linux
Linux over SSH
packages · services
files · users
scheduled tasks
VMware vSphere
Proxmox
clone · snapshot
full-state restore
PostgreSQL · MySQL
backup · restore
migrations
replication health
Docker
AWS · Azure · GCP
Slack · Jira
ServiceNow
Deploy
WeftRoute ships as a self-contained on-premises appliance and runs against a model on your own GPU — or against none at all. Nothing about your network, your configs or your incidents leaves the building. Signed images, SBOMs, and an operator CLI that verifies what it installs.
The appliance needs no outbound connection to function. Models are shipped in, verified and upgraded through the same signed-bundle path as the software itself — and rolled back the same way when a new one reasons worse than the old one.
Local model · no telemetry · no vendor cloud between you and your devices
Get started
That's enough for a first loop — a plan, an execution, a verification, and a cause.