Praxis.
The evidence-native development platform that learns a business and ships its software. Point Praxis at a company: it reads the docs, maps the data, builds a governed knowledge fabric and graph, stands up grounded agents, and runs a software factory that delivers real, reviewable code. Nothing Praxis claims, retrieves, or builds is ungrounded.
Perceive, prove, reason, produce.
Every unit of work runs the same loop. The Master conducts every step; Structura bounds the loop as its evidence spine. Agents answer from evidence packets with citations — or they abstain. There is no third option.
The Master conducts
A single orchestrator sequences every phase — nothing runs outside the loop, nothing skips a proof.
Structura bounds the loop
The evidence spine supplies governed packets with locators and citations; every claim traces to its source.
Cite or abstain
Agents answer only from evidence packets. When the evidence doesn't support an answer, they say so.
It learns the business first.
Before Praxis writes a line of code, it builds the knowledge to justify it: a Product KB for what the business does, per-source KBs for each system it reads, and a Master index that fuses them — with a knowledge graph so retrieval can follow relationships, not just similarity. Grounded agents draw from this fabric and nothing else.
- Product KB — the business, its domain, its language
- Per-source KBs — one per ingested system, independently governed
- Master index — fused retrieval across the whole fabric
- Knowledge graph — entities and relationships, traversable at query time
Intent in, audited software out.
Five stages, one command: praxis run-the-line. When downstream reality drifts from the spec, a back-propagation loop pushes the change upstream so the spec, blueprint, and work orders never lie. Watch the line run:
Evals as numbers, not vibes.
First measured eval run, local Tier A (Ollama nomic-embed-text + qwen3:8b) — from the repo's own eval reports, baseline included.
+0.152 recall over the keyless baseline on the same corpus and questions.
+0.255 groundedness over baseline; citation faithfulness reported honestly — it's the number being driven up next.
What the platform guarantees.
Self-building knowledge
Point it at a company and it builds Product, per-source, and Master knowledge bases on its own — a governed fabric, not a pile of embeddings.
Retrieval that follows relationships
A knowledge graph over the fabric lets retrieval traverse entities and dependencies, not just match similar text.
Intent to shipped, traced software
The factory turns intent into specs, blueprints, work orders, and code — every artifact linked back to the requirement that demanded it.
Prove it before you trust it
Evals as numbers, not vibes: recall, groundedness, and citation faithfulness are measured on every run and reported with their config.
Any provider, routed per function
The Model Hub routes each function to the right model — vLLM, Claude, or Ollama — swappable without touching the callers.
Give any agent the Praxis tools
praxis-flow init wires any agent runtime into the platform over MCP — 40 tools, every call policy-enforced.
A hundred agents, one live board
An SSE-backed claim queue coordinates swarms of workers on a single live board — claims, progress, and completions in real time.
See and cap every token
A usage ledger meters every model call; budgets cap spend per team, per project, per function — before the bill arrives.
Identity gated everywhere
SSO plus AWS-IAM-style policies on every route and tool: default-deny, deny-overrides, no unauthenticated side doors.
Default-deny, deny-overrides.
Identity and policy are the platform's floor, not a feature. Every route, every tool, every agent call passes the same gate.
- authorize()One authorization path everywhere — every route and MCP tool passes the same policy engine
- Keycloak SSOBrokered single sign-on across 9 identity providers
- Audit logAppend-only record of every privileged action
- Audit bundlesContent-addressed, replayable bundles — rerun the exact decision from its inputs
- PHI/PII shieldSensitive fields masked and HMAC-tokenized before anything reaches an index
Honesty note: compliance language in the repo describes target posture and the controls built to support it — not a completed certification.
“The future of enterprise development is not a faster way to write unverifiable code. It is a platform that knows the business, proves what it builds, and ships it under review.”