ownerlmnodes
    [ give your agent a node ]

    Private AI tools now. Property-hosted compute next.

    Connect agents to a portfolio's private context today. For work that needs to sit nearby, reserve limited-pilot GPU capacity inside properties that already carry the power, secure space, and cooling.

    install the mcp serverread the docs
    [ install the mcp server ]
    One line into your client. Streamable HTTP, JSON-RPC 2.0. Clients that speak OAuth will ask you to sign in and approve once, no key to copy. Anything else can use a bearer key from Settings inside any OwnerLM workspace.
    $ claude mcp add --transport http ownerlm https://ownerlm.com/api/public/mcp

    First call opens a sign-in and one approval screen. After that the agent works as you. Using Claude in the browser? Settings, Connectors, Add custom connector, then paste the server URL.

    [ what you can do today ]

    Live now, and honest about the rest

    Two of these are running. Two are at reservation stage. We would rather say so than sell a queue that does not exist.
    [ live ]

    MCP tools

    Around forty tools over Streamable HTTP: portfolios, properties, documents with cited semantic search, deals, vendor contracts, reports, skills, and agent runs. Bring a key and connect.

    [ live ]

    The harness

    Posture rules on what an agent may do, approvals on the writes that need a human, a full action ledger, and one-click undo from before and after snapshots. Planned to open source.

    [ reserving ]

    Edge GPU nodes

    Pilot racks inside real properties. Reservations are answered by a person, usually the same day. No self-serve provisioning API yet, and we will not pretend otherwise.

    [ reserving ]

    Site capacity

    Owners with power and space bring the host side. Sites under evaluation and their phased capacity show up on the compute market.

    [ two tracks ]

    Who this is built for

    Any model or agent is welcome to find us and reserve. These two are the ones we can serve unusually well.
    track 01

    Real estate and proptech AI

    Agents that work leases, financials, work orders, and whole portfolios. Reach our tools over MCP today, and reserve hardware near the properties and people the model serves.

    • *Live MCP server with portfolio, document, deal, and agent tools
    • *Semantic search over an owner's own documents, with citations
    • *The harness behind it: posture rules, approvals, and a full action ledger
    • *Planned to open source, so you can run the same agents on your own capacity
    track 02

    Edge inference and IoT

    Work that is bound by distance rather than by model size. Cameras, access control, property systems, and small models kept resident next to the sensors that feed them.

    • *Capacity in the metro your users and devices are already in
    • *Steady streaming inference instead of bursty training runs
    • *A known street address for data residency questions
    • *Private uplink into the property network when the workload needs it
    [ why the edge ]

    Why capacity here beats capacity far away

    Distance is the part you cannot prompt away

    A round trip to a distant region is fixed cost in wall clock. Inference in the metro answers faster than any tuning pass.

    Cost per token follows the power bill

    Most of a cloud invoice is markup on power, space, and cooling. We buy those at property economics.

    A street address beats a region code

    Regulated tenants and finance teams ask where the prompts and the weights sit. Here that has a plain answer.

    Racked in months, not build cycles

    Capacity is limited, but it is real, and it lands inside properties that already carry the service.

    [ hardware ]

    What goes in the rack

    AMD and NVIDIA accelerators matched to the workload rather than a house favorite. Same catalog our site configurator sizes against, so the specs here cannot drift.
    AMD
    Instinct MI210
    VRAM
    64GB
    TDP
    300W
    Form
    PCIe card

    Single-slot-friendly inference workhorse. Fits a normal server chassis in a normal building.

    PCIe 4 x16 · Passive, needs server airflow

    AMD
    Instinct MI250
    VRAM
    128GB
    TDP
    500W
    Form
    OAM module

    Large-memory module for bigger models kept resident in VRAM.

    PCIe 4 x16 · Passive, needs server airflow

    AMD
    Instinct MI250X
    VRAM
    128GB
    TDP
    560W
    Form
    OAM module

    The widest MI200 configuration when memory bandwidth is the limit.

    PCIe 4 x16 · Passive, needs server airflow

    AMD
    Instinct MI300X
    VRAM
    192GB
    TDP
    750W
    Form
    OAM module

    192GB per module for frontier-size models without sharding.

    PCIe 5 x16 · Passive, needs server airflow

    NVIDIA
    L40S
    VRAM
    48GB
    TDP
    350W
    Form
    PCIe card

    Mixed inference and rendering when CUDA is a hard requirement.

    PCIe 4 x16 · Passive, needs server airflow

    NVIDIA
    A100 80GB PCIe
    VRAM
    80GB
    TDP
    300W
    Form
    PCIe card

    Proven inference platform with broad framework support.

    PCIe 4 x16 · Passive, needs server airflow

    NVIDIA
    H100 PCIe
    VRAM
    80GB
    TDP
    350W
    Form
    PCIe card

    Top-end throughput where budget follows latency.

    PCIe 5 x16 · Passive, needs server airflow

    [ reserve ]

    Tell us what you need to run

    A person reads this, usually the same day. No sales sequence, no gated deck.
    track:
    [ pricing and terms ]
    No public rate card. A pilot is quoted per site, because the number follows the property's power, install, and term. Pilot terms are short on purpose so both sides see real numbers before a longer build.

    Prefer a phone? Call or text 816.300.7645.

    Reading this as a machine? The plain text version:

    https://ownerlm.com/nodes.txt