Harness Basic Architecture

One template. Two doors in. Every solution is yours to keep.

Every di-factory solution — from an open, ready-to-use build on our OpenClaw + Paperclip stack, all the way to a fully tailored one — sits on claude_harness, the same MIT-licensed template.

Eighty percent is the template we maintain in the open. Twenty percent is your instance: your prompts, your data connectors, your policies.

You own the code. You own the cloud. You own the data.

Two doors in

Door 1 — Open, ready to use. The full OpenClaw + Paperclip stack, sitting on the same template. Fastest path in, lowest lift, calibrated for teams who want the template's opinion on how AI should run.

Door 2 — Tailored. Same 80/20 discipline, calibrated to your workflow, your data, your policies. Slower to build, closer to fit.

Same team. Same template. Same ownership sentence at the end. What changes is how much of the twenty percent we shape together.

Why it's built this way

You own the system, not a subscription. No SaaS trap. No metered access. When we finish, the repo, the cloud account, and the data are yours. If you stop calling us tomorrow, the system keeps running.

No lock-in — swap models, providers, clouds. The template is designed to be provider-neutral. Change your LLM provider, your vector DB, your cloud — without a rewrite. Everything replaceable is replaceable at the config layer.

Zero markup on third-party costs. You pay OpenAI, AWS, Anthropic, Pinecone, whoever — direct, at their price. We don't resell compute or tokens. Our fee is our fee.

Built for Mexico first. PII tokenization, LFPDPPP-compliant baseline, Spanish-first prompts and evals when you need them. Not a US template with a translation layer.

1. What this is

Every solution di-factory delivers — a tailored ML model, an MVP, an agentic transformation, a PyME automation — is built as an instance of claude_harness, our open-source (MIT) solution template. 80% is the template: the agent core, tool registry, memory, governance and ops layers we build and maintain in the open. 20% is your instance: the solution spec that encodes your prompts, your tools, your data connectors and your policies. Onboarding a client means writing that spec and connecting secrets — it does not mean writing new code.

Every di-factory solution is tailored — from an open, ready-to-use build on our own OpenClaw + Paperclip stack, to a fully custom one. Same template underneath. Same code, cloud, and data ownership at the end.

The template is open source and MIT-licensed: it runs in your own cloud account, on your keys, under your control. There is no lock-in to one model, one cloud, or one runtime — and third-party costs (model calls, messaging gateways, cloud infrastructure) pass through at zero markup, itemized per vendor.

The template lives in the open on GitHub, MIT-licensed, versioned, and audited: github.com/di-factory/claude_harness. Read the code before you sign anything — that's the point.

2. What's in the 80% — capability modules by line of business

The module matrix below shows which parts of the template a given instance actually needs. It is the same evidence base every line of business draws from — nothing here is bespoke per client.

● required · ○ optional · – not needed

3. What's in the 20% — the solution spec

A solution is a versioned, declarative spec — JSON, validated by schema — with prompt files and optional extensions alongside it. This is the file format your 20% gets written in:

A pack is a reusable, versioned spec fragment — for example a "PyME Appointment Agent" base. An instance is one or more packs plus your overrides plus a tenant id. For a clinic's Appointment Agent, the 20% is the reminder templates, the Google Calendar connector, and the clinic's own patient-list rules — everything else is template.

4. How it's different

5. Roadmap

The diagram above tracks the actual build — five 4-week milestones from 2026-09-28: core skeleton (M0, 2026-10-23), agent core and console (M1, 2026-11-20), headless runtime and governance minimum (M2, 2026-12-18), intelligence modules (M3, 2027-01-15), and v1.0 operations hardening (M4, 2027-02-12). What this page claims publicly is gated to what has actually shipped and passed its tests at each milestone — not the roadmap date alone.

5.1 Governance

Shipping in M2 (2026-12-18). PII tokenization, consent and opt-out tracking, an immutable audit log, and per-data-class retention — built for Mexico first against the LFPDPPP baseline. This section will expand once the governance modules ship and pass their acceptance tests; nothing here is available before then.

6. Not an eighth line of business

The template sits underneath di-factory's seven lines of business — it is the delivery mechanism, not a competing offer. Tailored ML Models, MVP Development, Business Consulting, VC Partnership, Agentic Transformation, OPC Startup and SMB/PyME Solutions are all things you buy from di-factory; an instance of this template is how we build the one you choose.

Want to see the template running? Book a 30-minute discovery call — we'll show you an instance and walk the 80/20 split for your use case.
Book a discovery call →