Open Source·May 2026·Prism
One package today. More to come, in the order they prove out in production. Below: what we publish, what we keep, and why the boundary lives where it lives.
I·Published Packages
Phase 8.6 shipped the customization engine to npm in May 2026. The package is the orchestration framework underneath every Prism build, the LLM call wrapper, the validator, the parsers, the safety guardrails. It is already running in production against thousands of builds.
@prism/customization-enginev0.1.0
The orchestration framework underneath every Prism build.
The pure pieces, the LLM call wrapper, the BuildPlan validator, the file-block parser, the safety validators, the never-throws contract, published as a portable package. Around 2,000 lines of TypeScript, 30 named exports, zero runtime dependencies on Prism. The proprietary parts (the 17-pattern domain analyzer, the prompts themselves, the Supabase and GitHub glue) stay in-tree. The package is the substrate vocabulary; the in-tree code is the company.
II·What We Publish, What We Keep
We open the framework. We keep the content. A package consumer gets the orchestration shape, the never-throws contract, the JSON-block parser, the safety validators, the LLM call wrapper. They do not get our prompts, our 17-pattern domain analyzer, our integration matrix, or the schema-evolution rules. Those are the part of the substrate that took years to tune.
The point of publishing the framework is not to give away Prism. It is to lock the vocabulary. When an engineer six months from now reaches for an LLM-driven code-customization framework, the words and the shape should already be ours.
III·Open Source Tenets
Vocabulary leadership beats vocabulary defense.
MCP became the industry standard because Anthropic published the spec, shipped working clients, and donated it to the Linux Foundation before the market settled. We are not Anthropic, and the customization engine is not MCP, but the posture is the one to copy. Publish the words. Run them in production. Let the rest of the industry adopt them.
The boundary is sharp, on purpose.
We publish the WHAT-engine, the framework that any LLM-driven code-customization project would need. We keep the HOW-Prism, the specific prompts, the domain analyzer, the integration matrix, the schema-evolution rules. The package gives an engineer enough to build their own orchestrator; it does not give a competitor the Prism product.
Open source is a substrate move, not a marketing move.
Every package we publish is something we already use in production. Nothing is open-sourced as a demo. The customization engine has shipped against thousands of builds before it landed on npm; the orchestration framework was a thing before the publish, not because of it.
IV·What's Next
Candidates exist. The npm-registry probe + format-preserving package.json rewriter (the deterministic core of the Atlas Janitor) is one. The schema-evolution validator with its 21 hard rules is another. The MCP tool registry shape is a third.
None of them are scheduled. We publish a piece when it has shipped, stabilized, and earned its place outside our walls. The list above is the parking lot, not the commitment.
Read the ManifestoMay 2026 · Prism · Open Source