We install open-source AI on your own computers, shape it around your recurring work with the TresPies Harness, and stay on call to keep it useful. Built for nonprofits and small offices that handle records they can't send to someone else's servers.
Each install follows the same six steps. The first five are setup. The sixth is the retainer, and we expect it to be where most of the value comes from, because the useful routines come from watching how your staff actually use the tools.
Short intake sessions with the teams who'll use it. We list the work that repeats every week or month, and we map your data: what's routine, what's sensitive, and what must never leave the building.
We look at the computers you have. The Harness runs on Windows 10 or 11, and installing it needs administrator access, which we arrange with your IT person. If none of your computers can run a capable model at a usable speed, we write a spec for one shared workstation that you buy directly, at cost, from any vendor you choose. We pick the open model to fit the machine.
OpenCode, a local model runtime, and the TresPies Harness go onto your machines. Private mode and the guardrails are on before anyone types a word. No commercial AI service is installed. If you later want one for non-sensitive work, it is a separate setup, added only on your written decision.
Three starter routines for your most common tasks, tested on real examples you pick from non-sensitive files. Each one comes with a two-page quick guide.
Two hands-on sessions, in English or Spanish as your staff prefer. People work through their own tasks, not a demo.
One working session a month with whoever is using it. Email questions get a first reply within two business days. Up to two new routines each quarter, chosen by you. Every quarter we update and test the software and the model, and you get a one-page record of what changed and what's next.
The engine is free, open-source software. The model runs on your hardware. The Harness is our layer on top: the setup that turns a general AI agent into something an office can trust with its daily work.
Named assistants, ready-made routines, guardrails, a private mode, house rules in your voice, and a session log. Licensed to you while the retainer runs. If you stop, you keep the version you have, without further updates, including security fixes.
A widely used open-source AI agent that reads, organizes, and writes files in a working folder, and works with models that run entirely on a local machine. Yours to keep using with or without us. opencode.ai
Your computers, running an open-weight model, used under its own license, through a local runtime such as Ollama. No per-seat AI subscription is required.
A drafter, a records clerk, a translation drafter, and a reviewer. Each has one job, its own instructions, and its own permissions, so staff open the assistant for the job and it has no more access than that job needs.
Written playbooks for your recurring tasks: a monthly report section, a resident notice in two languages, a spreadsheet cleanup. Staff run them by name instead of writing prompts from scratch.
Assistants can read the working folder you give them. Each change to a file is shown and waits for a person's approval, and changes stay in its Drafts folder. A person can switch this off for one window, by turning on auto-approve or allowing all of that window's changes; the log records each reply. The assistants have no tools to run commands or programs, or to browse the web.
Only the local model is available, and web access is denied. It is the only mode the first version installs.
A plain-language rules file with your organization's style, glossary, preferred Spanish terms, and a list of things the assistants are instructed never to do.
What was asked, what was produced, and how each change request was answered, kept on your machine. Attached files are named, not copied. OpenCode also keeps its own conversation history on the computer, including the text of files the assistants read, until it is cleared. We agree the schedule with you at install, and only a full clear removes it completely. Treat both like your other sensitive records.
// Harness private mode (excerpt): machine settings plus one assistant's own rules { "$schema": "https://opencode.ai/config.json", "share": "disabled", "enabled_providers": ["ollama"], "provider": { "ollama": { "npm": "@ai-sdk/openai-compatible", "name": "Office model (local)", "options": { "baseURL": "http://127.0.0.1:11435/v1" } } }, "permission": { "*": "ask", "read": "allow", "glob": "allow", "grep": "allow", "list": "allow", "edit": "ask", "bash": "deny", "webfetch": "deny", "websearch": "deny", "external_directory": "deny" }, "agent": { "records-clerk": { "description": "Cleans spreadsheet copies; shows every change first", "mode": "primary", "permission": { "edit": { "*": "deny", "*/Drafts-Borradores/*.csv": "ask", "*/Drafts-Borradores/*.md": "ask" } } } } }
The guardrails are ordinary OpenCode settings, plus the launch settings and network rules we install, all written down for each computer, so your IT person (or anyone you hire later) can read exactly what the assistants are allowed to do.
Every engagement starts with the same setup and quarterly retainer. Add-ons are scoped in a one-paragraph note (what, by when, how it's billed) before any work starts.
Pricing is set in the first conversation, by size and scope, with nonprofit rates.
Open models running on one office machine are slower and less capable than the largest commercial services. That's the trade for keeping the AI, and the work you give it, on your own machines. We set expectations to match.
Private mode does not cover Windows' own services (error reports, typing samples, cloud clipboard) or the network-name lookup a document can trigger, which no setting fully closes. Your IT person decides the Windows settings, and we list them for each computer.
Bring the three tasks that eat the most of your team's week and a sense of which computers you have. We'll tell you plainly whether local AI will help, and what the first month would look like.