How it works
Compiling a repository into a Company Brain.
Praxis, the compiler behind ELDRION, turns a repository into a organizational digital twin in four stages: it compiles the repo into structured evidence, cites every claim back to its source, certifies executable skills against human approval, and serves the result to AI agents over MCP. The compile stage runs entirely on your own machine.
How does the pipeline work?
Repository to evidence
The compiler scans the repository, commit history, decision records, and ownership maps, and extracts them into a structured knowledge graph. This stage is deterministic and offline: it uses no LLM, so the same repository state always produces the same output.
Evidence to claims
Evidence becomes claims, and every claim carries a permalink to the exact source it was derived from. Nothing in the brain is an assertion you have to take on trust — each one resolves to a verbatim excerpt of your own repository.
Claims to certified skills
Claims are drafted into executable skills — rules an agent can act on. A draft becomes servable only when a named human on your team certifies it, and the sealed artifact carries a content hash so a served skill is provably the one that was approved.
Skills to your agents
The brain runs as an MCP server. Claude Code, Cursor, Codex, and Gemini query it directly, so organizational context reaches the agent before code is written rather than during review.
Why is the build deterministic?
Because a claim you cannot audit is not evidence. It would be easier to build this with a model in the loop — point an LLM at a repository and ask it to summarize the architecture — but that produces something that reads well and cannot be checked. You cannot tell which parts were derived and which were generated, and you get a different answer next week from the same input.
Compiling instead of generating is what makes the output checkable. Every claim traces to an artifact, every skill traces to the person who approved it, and a rebuild on unchanged input produces an unchanged brain. That property is also what allows the compile stage to run locally with no network dependency at all.
What leaves your machine?
The build is local: your source stays on your disk, and no third-party service sees it. What can leave — only when you explicitly ask — is the sealed brain artifact: cited claims and certified skills, not raw code. Publishing it to a gateway is how a team serves the same certified skills to everyone, and the gateway is self-hostable if you would rather it never crossed your own infrastructure.
Operational telemetry is stricter still — a few local counters written to your own disk, with no network path at all.
Common questions
Does my source code leave my machine?
No. The build scans the repository on your own disk and produces a sealed artifact of cited claims and certified skills. Raw source is never uploaded. Publishing that artifact so a team can share it is a separate, explicit command.
Does the compiler use an LLM?
No. The build stage is deterministic and offline by design, which is what makes the output auditable: a claim traces to the artifact it came from rather than to a model's generation.
Which AI clients can consume it?
Any MCP client. Claude Code, Cursor, Codex, and Gemini are what design partners use today.
Can I use it yet?
Praxis is in private beta with design partners. Access is by invitation — write to contact@eldrion.in if you want to be considered.