PERSONAL PROJECT · SELF-HOSTED · LOCAL-FIRST

One operating layer
over the systems I use.

Franklin is a human-governed AI operations environment I built part-time to bring infrastructure, maintenance, workflows, information and local AI into one place.

OPERATING PRINCIPLEAutomate everything except decisions.
Click any screenshot to open it full size.
Franklin overview showing decisions, systems report, plugins and local AI panel
BUILTPart-time
MODELLocal-first
CONTROLHuman-governed
STATUSOperational
RELEASEPrivate
WHY I BUILT IT

I didn't want another dashboard.

My home environment had become a collection of separate systems: virtualization, containers, storage, maintenance, media, news, workflows and AI. Each worked, but operating them meant jumping between interfaces and rebuilding context in my head.

Franklin is the layer above them. It gathers evidence, keeps state, runs safe workflows, surfaces decisions and gives me one conversational interface to the whole environment.

The current build is a private personal deployment, not a general-purpose product or public release.

ARCHITECTURE

The model is not
the control plane.

Franklin separates conversation from authority. AI helps interpret state and prepare work; evidence, capabilities, workflows and approval boundaries determine what can actually happen.

AUTHORITY You judgment · approval · decisions
OPERATIONS LAYER Franklin
ConversationDecisionsWorkflowsReportsMemory
CAPABILITY + EVIDENCE LAYER
PluginsFresh stateChecksBounded actionsVerification
ProxmoxDockerTrueNASLinuxPlexARRNewsRedditWeatherFiles
REASONINGLocal AIOllama · local models
ENGINEERINGCodexscoped remediation · future loop
THE SYSTEM

Conversation backed by evidence.

Franklin is useful because the chat is not supposed to be the source of truth. The operational layer is.

01

Ask what needs attention.

Franklin turns system state into an operator-facing conversation, prioritizes what matters, and carries evidence quality into the answer instead of silently treating unknown state as healthy.

Franklin conversation answering what needs attention and explaining evidence quality
02

See the evidence.

Maintenance combines targets, findings, security updates, reboot state, provider checks and evidence freshness.

Franklin maintenance view
03

Automate repeatable work.

Scheduled and manual workflows refresh infrastructure evidence, run health checks and perform bounded operational tasks.

Franklin workflow list
04

Operate real infrastructure.

Plugins expose systems such as Proxmox through a consistent interface while keeping consequential actions separate from ordinary read-only inspection.

Franklin Proxmox plugin showing virtual machines, resources, storage and tasks
DESIGN PRINCIPLES

AI can reason.
It doesn't get to invent reality.

Franklin is being designed around a simple separation: systems provide evidence, AI helps interpret it, workflows perform bounded work, and the human remains the authority.

01

Evidence before action

Fresh system evidence should support operational conclusions and remediation.

02

Human approval

Consequential changes stop at a decision boundary instead of disappearing into autonomous execution.

03

Least privilege

Capabilities are constrained to the access needed for the job rather than handing an AI unrestricted control.

04

Verify afterwards

Execution is not success. The system should independently check the resulting state.

BEYOND INFRASTRUCTURE

One environment, not one use case.

The same plugin model also brings personal information and media workflows into Franklin without turning them into separate applications.

Discovery & media

Browse popular media, check availability and hand items into the media stack from the same environment.

Franklin media discovery plugin

Information

News, interests, Reddit and weather live beside operational tools rather than in another pile of tabs.

Franklin news plugin
CURRENT REACH

Different systems.
One operating environment.

Franklin's breadth comes from plugins rather than pretending every system is the same. Each integration contributes the capabilities and evidence it can actually support.

Proxmoxvirtualization
Dockercontainers
TrueNASstorage
Linuxmaintenance
Plexmedia
ARRautomation
Newsinformation
Redditinformation
Weatherlocal context
Filesworkspace
THE STORY

Built part-time by a tractor-parts guy.

I'm not a software company or an SRE team. My day job is in tractor parts. Franklin grew out of a practical question: why am I still doing all of this manually?

I used AI-assisted development heavily, but the architecture, requirements, operating rules, testing decisions and judgment stayed mine. Over roughly a year, the project grew from an idea into software I now use in my own environment — including remotely during normal day-to-day use.

ObserveGather authoritative state.
InvestigateWork out what actually matters.
VerifyCheck evidence and freshness.
RecommendPrepare the next safe step.
Human decisionI approve consequential work.
ExecuteRun bounded actions or engineering work.
Verify againProve the resulting state.
CURRENT STATUS

Working software.
Still being sharpened.

Franklin is operational in my own environment, but this version is intentionally personal and is not currently offered as an installable public release.

The remaining work is increasingly about orchestration quality: selecting the right capability from natural language, drilling into the evidence behind an answer, and making sure the model never claims it checked something unless a real tool actually did.

I show the rough edges because they matter. A believable wrong answer from an AI operator is more dangerous than an obvious failure.

SOURCE & AVAILABILITY

Shown publicly.
Not packaged publicly.

Franklin V2 is deeply shaped around my own environment, workflows and operating rules. Publishing that code as though it were a ready-to-install product would be misleading.

For now I'm documenting the architecture and showing the working system. If there is genuine interest in running something like Franklin, a future public version should be designed deliberately around configuration, onboarding and safe defaults rather than carved out of my personal deployment.

WHERE IT'S GOING

From assistant to governed operator.

The next major step is a grounded engineering loop: Franklin detects and investigates a problem, prepares a scoped remediation for a coding agent, waits for human approval, then independently verifies the result. The goal isn't autonomous authority. It's removing the mechanical work around human judgment.

CONTACT

Want to talk
about Franklin?

Franklin is a personal project, but I'm happy to hear from people interested in the architecture, human-governed AI operations, homelab automation, or where the project is heading.

[email protected]