Open Core
Open-source the framework so anyone can build with it and own it; charge for the hosted, integrated, pre-configured version for everyone who'd rather not assemble it. The open core isn't a giveaway — it's distribution.
The model
Open-core splits a product into two layers:
- The core (open source, free). The framework/engine — MIT or Apache, self-hostable. The user owns what they build, no lock-in. This is the credibility and distribution layer.
- The shell (paid, hosted). Deployment, integrations, configuration, governance, observation, and support — the assembly a buyer would otherwise do themselves. This is the revenue layer.
You make money not by gating the core capability but by selling the convenience and completeness around it: one-click hosting, integrations already wired, configs already chosen, the compliance/SSO/audit a team needs, and someone to call.
Why it works
The apparent paradox — "if it's open source, why will anyone pay?" — dissolves once you see the open core as distribution, not a giveaway:
- Builders who'd assemble it themselves can — and they become advocates, contributors, and the top of the funnel. (This is Build, Don't Buy from the user's side: build it yourself if that's your thing.)
- Everyone who'd rather not pays to skip the assembly. The afternoon of setup — wiring integrations, choosing configs, standing up hosting — is exactly what the hosted tier removes.
- Trust and no lock-in lower the barrier to adopt. Open code is auditable and portable, so the buyer commits more easily; the real switching cost accrues later in the hosted data, observation, and configuration.
- The open core is the cheapest distribution channel there is — GitHub, Show HN, integrations, ecosystem contributions (Distribution Is the Moat).
Examples
| Open source (build it yourself) | Paid (hosted, wired in) | |
|---|---|---|
| Reflex | Apache-2.0 Python web framework (pip install reflex, "you own everything you build") |
Reflex Cloud/Enterprise — deploy + Snowflake/Databricks/Okta integrations + SSO/SOC2 |
| private page | MIT persistent-agent framework / control plane | Hosted spin-up + 161 skills / 39 MCP integrations + presets/configs + gateway + observation + fleet |
| Loop | OSS operator desk for agent skills | Operator tier ($5/mo) — unlimited automations, model pick, marketplace payouts |
| Lineage | Vercel (Next.js), Supabase, GitLab, HashiCorp, Grafana | hosted / enterprise tiers atop an open core |
private page
Kevin's portfolio instantiates the model directly: the projects he builds to avoid buying SaaS (Build, Don't Buy) ship as open frameworks and monetize the hosted assembly. Sigil UI is MIT today (open core without a paid tier yet); Loop and private page already pair an open core with a paid hosted tier.
Open-core vs services-as-software
Different monetization shapes, often confused:
- Open-core sells the assembled tool — you still operate it (deploy the worker, run the app). Margin is hosting + integrations + governance.
- Services-as-Software sells the outcome/labor — the agent does the job; you buy the result. Margin widens as models improve.
They compose: private page is open-core today (you operate the worker), and its endgame — agents that deliver outcomes — edges toward services-as-software. Open-core funds distribution now; outcome-pricing is the deeper margin later.
Anti-patterns
- Gating the core capability so the "open" version is a crippled demo — kills trust and the distribution flywheel.
- Open-sourcing the differentiated/compounding layer instead of the commodity framework — give away the framework, keep the hosted data/observation/registry stickiness.
- Charging for what's trivial to self-host rather than for the assembly people actually don't want to do.
- Treating open source as a marketing checkbox without a hosted product that captures the value.
Concept Position
| Field | Value |
|---|---|
| Concept family | Agent harness and runtime primitives |
| Concept owned | Open-source the framework so anyone can build with it and own it; charge for the hosted, integrated, pre-configured version for everyone wh... |
| Category map | Concept System Map |
Clean-room reimplementation pressure
PhotoCraft is an open-source, clean-room reimplementation of Photoshop in Rust from the ArtCraft project, which has done the same for other Adobe apps. Eric S. Raymond reads it as the end of closed source. The account of how it was built (decompile, convert to a spec, generate Rust) is Raymond's guess, not the project's statement. Source: X/@esrtweet thread, 2026-10-06; Source: storytold/photocraft
The durable point for open core: when agents can rebuild a known product's surface cheaply, the moat is not the code. It is distribution, data, the hosted service and the speed of the team. Version checked: photocraft v0.20.1 (2026-11-03), Apache-2.0, early alpha. Source: gh release list, 2026-11-03
Mitchell Hashimoto's "The Building Block Economy": the most effective path to adoption is no longer a polished mainline app but building blocks that let others build, quantity over quality, as Ghostty's library spread showed. Source: X/@mitchellh, 2026-04-07
The same view from the other side: Brett Goldstein holds that 99% of software can be built with the same finite elements, so rebuilding them from scratch each time wastes tokens. Source: X/@thatguybg, 2026-07-18 Garry Tan quote-posted it as a pitch for GSkills. Source: X/@garrytan, 2026-07-19
Timeline
- 2026-07-01 | Concepts category refresh added this page to the Agent harness and runtime primitives family, linked it to Concept System Map, and kept it standalone because it owns this reusable mental model: Open-source the framework so anyone can build with it and own it; charge for the hosted, integrated, pre-configured version for everyone wh... private page
- 2026-06-15 | Page created from Kevin's framing linking Build, Don't Buy to private page: open-source the persistent-agent framework, charge for hosted machines/agents with integrations + configs built in — the Reflex model. private page