Skip to main content
Builder-grade. Self-hosted. Local + cloud.

Hermes Starter: Local + Cloud AI Deal Agents

An installable agent runner that works a real estate deal end to end, source to manage, on models you run. The private work stays on a local model on your own machine. Only the hard parts escalate to the cloud, and anything sensitive is local-only and fails closed.

Built for technical operators who want to self-host. No paid API key to run the local path. Pure Python, nothing to install. This is the starter version of the always-on fleet we build for clients, not a copy of it.

Download the pack

Want the system, not just the files?

Get the full pack once, then use the walkthrough below to see exactly what is inside and how to run it.

Get Hermes Starter

Enter your details and the full kit downloads right away. No cost, no paid key to run it.

We will email you occasionally about real estate AI. Unsubscribe anytime.

Why local + cloud, and why now

The resilient move is to own your stack: keep the cheap, private work local, escalate only the hard parts, and never have a single point of failure. As of 2026 the local models are finally good enough to make that real for a scoped deal agent.

The private work stays home

Sourcing, screening, underwriting, diligence, asset management. The cheap, high-volume 60 to 70 percent of a deal team's work runs on a local model on hardware you already own. No tokens billed, no data leaving the building.

The cloud is a booster, not a dependency

Only the hard synthesis, the IC memo, or an oversized context escalates to a frontier model, and only when you opt in. One config line moves any agent between local and cloud.

Fail closed on anything sensitive

Argus and its tenant data are local-only. If the local model is down, the runner refuses to fall back to the cloud and fails closed. That is the privacy guarantee, demonstrable with no key.

Size stopped predicting capability

A 3GB local model now handles scoped agent work that used to need a datacenter. The kit ships a verified, honest model matrix so you run the right model on the hardware you have.

The seven agents

The full deal lifecycle, reframed as run them yourself, 24/7. Each agent ships a drop-in system prompt and, where there is math, a deterministic Python script so the numbers are correct even on a tiny local model.

Sage · source

Scores every inbound deal against your mandate and returns kill, watch, or pursue. Runs 24/7 so only the deals worth a human reach you.

Forge · screen

Rapid go/no-go on the value-add spread, stabilized yield-on-cost over the exit cap, plus the early structural flags.

Atlas · underwrite

Full underwrite, NOI to five-year IRR, and the exact price that clears your hurdle. The math is real Python, not a guessed token stream.

Echo · market

Loss-to-lease, comp position, and the supply pipeline read, separating real embedded upside from hopeful upside.

Argus · diligence

Scans the rent roll and T-12 for red flags. Local-only by default, because tenant data should never leave your machine.

Rex · memo

Synthesizes the team into one IC recommendation, with the bid that clears the hurdle and the conditions to clear first.

Ledger · manage

Trailing-12 NOI variance against budget, called plainly before it surprises an LP.

The local to cloud router

Each agent is assigned an endpoint in one config block. Switch a model without touching code. Mark anything sensitive local-only and it fails closed instead of quietly leaving the machine.

config.yaml

agents:
  atlas:                      # underwriter
    route: local
    fallback: [local, cloud]  # escalate only if local fails
  argus:                      # diligence, reads tenant PII
    route: local
    local_only: true          # NEVER falls back to cloud
  rex:                        # the one hard-synthesis step
    route: cloud
    fallback: [cloud, local]

Run the sample deal end to end with no key:

python3 -m hermes.run demo        # full pipeline, bundled mock, zero keys
python3 -m hermes.run route-demo  # see the three routing branches

Then install Ollama or LM Studio, point one config line at it, and the same pipeline runs on a real local model. The verified model matrix tells you which model fits your hardware.

Questions real estate teams ask

What is a local AI deal agent for real estate?

A local AI deal agent runs on a model hosted on your own machine, using a tool like LM Studio or Ollama, rather than sending data to a cloud API. For a real estate operator, this means the sensitive stages of a deal (rent roll analysis, tenant PII, diligence documents) stay on your hardware and fail closed if the local model is unavailable, instead of quietly uploading to a third-party server.

How do local and cloud AI agents work together for CRE deal teams?

A model router assigns each agent either local, cloud, or local-with-cloud-fallback in a single config line. High-volume, private work (deal sourcing, screening, diligence) runs on the local model. Hard synthesis steps that need a frontier model, like the IC memo that holds the whole team's findings, escalate to the cloud only for that task. Sensitive agents are configured local-only and fail closed rather than falling back.

Can a self-hosted AI agent run real estate underwriting calculations correctly?

Yes, when the math runs in code rather than in the model. Hermes Starter routes all financial calculations, NOI, cap rate, DSCR, cash-on-cash, five-year IRR, equity multiple, and the bid price that clears your return hurdle, through an auditable Python module. The local model reads the deal and calls the function; the numbers come from the code, not from the model's inference.

Where AI changes the answer

A local model now handles the private, high-volume 60-to-70 percent of a deal team's work — sourcing, screening, diligence — on hardware you already own, while escalating only the hard synthesis steps to a frontier cloud model, so client data never leaves the building and the cost of running an always-on deal team drops to a fraction of an all-cloud setup.

See how we build this →

Want the system, not just the files?

Two ways to go further with NextAutomation.

Work with us

Have us build the sourcing-to-reporting pipeline these files only hint at. Custom AI systems for your firm.

Apply to work with us

Join the community

Live builds, the template library, and a room of operators using AI on real deals. from $49/mo.

Join the community