workflow builder
Human-approval gates from intent to execution
OWNER · MIDDLE · SUB
A project in orchestration systems
Make room for a verdict.
xnchSystems is an optional attempt at governed autonomy: an agent that can perceive, remember, decide, and act — without presenting that judgement is free.
A model — or ten thousand of them — deciding whether an act is worth taking. If it is not, we do not have to guess.
02
Owned control planes on internal infrastructure
01
Binding, non-bypassable evaluation for every action
∞
Human veto — the decision ledger is append-only
The system
should know what
it is allowed to do.
Most agent demos skip from prompt to outcome. xnchSystems stays with the uncomfortable middle: policy, memory, an approval boundary, an audit trail, then learning.
This is not a product-shaped abstraction; it is a small, working system with a point of view about agency. The control plane is called xnch. The decision engine is called nexi. They are separate because conflating them is how you end up rewriting an audit trail to make a feature work.
Visit the architecture01 — architecture in brief
A boundary you
can point to.
Two services, one deliberate seam. Decisions can be routed through the system without becoming inseparable from it.
system topology — v1control / decision
XNCH · control plane
policy
Policy
check + recall
memory
Memory
read + recall
contract
Contract
assert type
ledger
Ledger
append only
owner · policy · sessiondecision ledgeroutput enforced
seam
NEXI · decision engine
directive
Directive
Step the plan
recall
Recall
Step past now
act
Act
Skip the ask
learn
Learn
Change the next
perceptionmodel policyexecution
Policy before power
The permission model evaluates intent before an agent gets a shell.
A human can say no
ANY planned action that wants a system action waits on a human before anything external happens.
Every action leaves a trace
The first ask of any part of the architecture: can an auditor see it? Not an afterthought.
02 — the operating loop
From signal
to learning.
Most model throughput is treated as a character goal, and reactive systems get more frantic behaviour. A single loop can make the next one smarter.
- 00PERCEIVERead the pull
- 01RECALLBring back what matters
- 02DECIDEChoose the next act
- 03CONSTRAINCheck model policy
- 04PROPOSEForm an intended act
- 05VERIFYAsk a channel model
- 06EXECUTERun the approved act
- 07OBSERVEWatch the outcome
- 08RECORDWrite the execution
- 09LEARNShape the behaviour
03 — things that exist
A studio,
not a feature list.
Every artifact is a single idea on the same problem. The work is in the state outside the code or in the work that goes into the code itself.
governed agents
Command and recall agents with different powers
LAST DECIDED BY ROUTER
MCP bridge
A controlled bridge to tools, not a tool free-for-all
CONNECT · 14 SERVERS
voice loop
A CLI voice loop that keeps the interface close to reasoning
VOICE · LOCALHOST
training + eval
Training and evaluation as a living feedback surface
V2 ROUTER · V2 TONE
runbooks + docs
Architecture docs that explain the trade-offs, not the slogans
ARCHIVE · V1 ARTIFACTS
04 — two machines
No k3s.
On purpose.
xnchSystems runs on two owned physical machines under a no-k3s system regime.
That choice is not a rejection of scale. It is a choice to keep the operational surface small enough for one builder to reason completely. More boring. More debuggable. A system whose failure modes can be audited.
machine A · mx680
deterministic · keepalive
machine B · SM710
decision engine · nexi
05 — field notes
Read the
receipts.
A public repo for a private system. The docs are the work, not an appendix.
Read the documents