Compile your repository into a cited Company Brain.
CLI / Studio is the developer execution surface for the Company Brain — it runs locally, compiles your repositories and team history into a deterministic, cited knowledge graph, and serves certified skills to Claude Code, Cursor, Codex, and Gemini over MCP. Your source never leaves your machine.
Compile
Repositories, ADRs, and team history are compiled locally into a deterministic knowledge graph. Raw source never leaves your machine.
Cite
Every claim is a verbatim, permalinked excerpt of its source — no paraphrasing, no invented conventions.
Certify
A named human reviews and approves drafted rules before they become an executable skill. Nothing ships unapproved.
Serve
Certified skills are queried over MCP by your AI tools — cited, and accountable to the person who approved them.
Convention correctness primary
Your agent stops guessing your team's style from the internet's average of every codebase. It pulls the rule your team actually wrote, not a plausible-sounding default.
An audit trail
Every rule an agent followed traces to a source permalink and the name of who approved it. An answer, not a shrug, when someone asks why it did that.
Survives turnover
The engineer who knew why the migration has to run before the deploy can leave. The certified rule doesn't — it's still in the served brain, still cited.
Faster ramp-up for whoever's next
The next engineer on the project isn't starting from a cold repo and a Slack search. praxis onboard <area> hands them the certified rules and the cited threads behind them — the tribal knowledge that used to take weeks to pick up by osmosis, on day one instead.
Fewer tokens per session side effect
Targeted retrieval beats an agent cold-reading the repo. Real, but it falls out of the ranking mechanism — it isn't why this exists.
What actually prevents hallucinated conventions: a claim's text must be an exact substring of its source. Checked at build time, not by another model's judgment. Paraphrase or invent a rule, and the build fails closed.
"...after the flaky CI run we settled on running ruff before merge, line length 88, update your pre-commit config by Friday."
"running ruff before merge, line length 88"
astral-sh/ruff#482 · approved_by: j.kesslerA step reading "the team prefers ruff" instead of the exact quote above never reaches certification — it isn't a substring of any source, so it's rejected before a human ever sees it.
Provenance ≠ correctness
A citation proves a human once said this and stands behind it — not that the rule is right for the code you're touching today. Tests, review, and CI still do the job they've always done.
Extraction quality doesn't transfer cleanly
The judge tier hits 1.000 precision at 0.778 recall on its hand-labeled holdout — but the earlier filter is repo-tuned: 0.750 precision on the repo it was tuned on, roughly 0.091 on one it wasn't. Lower precision costs review time; it never ships an unfaithful claim, because faithfulness is the substring check, not the filter's judgment.
A human can still certify a bad-fit rule
The gate stops fabrication, not poor judgment. Nothing stops a certifier approving a real quote that turns out to be the wrong call.
pip install --force-reinstall --extra-index-url https://eldrion.in/dist/simple/ praxis-cli praxis login # GitHub Device Flow — authorizes this device praxis init praxis build # compiles locally, deterministically praxis skills draft --docs . # draft skills from the compiled evidence praxis skills approve <skill> # you certify it — named approver praxis push # publishes the sealed brain — skills + cited claims, not source praxis connect # wires Claude Code, Cursor, Codex, Gemini over MCP
This is the real, working install command. CLI / Studio is currently in private beta — it requires an approved invite, and the command above won't resolve without one.