The multi-agent layer that learns with your team.

Personal → org memory loop

Senior staff's "magic prompts" auto-promote into shared knowledge via three independent paths: frequency, outcome correlation, and LLM self-eval.

3-path promotion engine

Frequency-based, outcome-correlated, LLM-scored. Run in parallel — never depending on a single signal. Configurable thresholds for auto-promote vs review.

Skills also promoted

Not just memory — your personal skills get tracked, scored, and promoted to the org skill catalog when they prove themselves.

Outcome tracking built-in

record_outcome() attaches success/failure to episodes. The consolidator uses these signals statistically — no separate analytics pipeline needed.

Memory mode toggle

Per-user switch: accumulate (default) or read_only. Read-only sessions silently drop writes — useful for sensitive content. Admins can lock the mode tenant-wide or by role.

Multi-LTM fusion + routing

Run several LTMs in parallel and fuse with Reciprocal Rank Fusion — or route per query (temporal → Zep, audit → JSON, entity → Mem0). English + Japanese keyword detection. Higher recall without picking a winner.

6 LTM backends

JSON, Mem0, LangMem, Letta, Zep, HindSight — switch with one line. Plus Graph layer (optional) for relationship-heavy domains. Zero vendor lock-in.

Multi-LLM (15+ first-class · 100+ via LiteLLM)

Claude, ChatGPT, Gemini, Gemma, Qwen-API, Qwen-local (Ollama), DeepSeek, Mistral, Grok, Llama, Cohere, Perplexity, Phi + 100+ via LiteLLM. Same models on enterprise clouds: azure/* (Azure OpenAI), azure_ai/* (AI Foundry), bedrock/* (AWS Bedrock), vertex_ai/* (GCP Vertex AI). Auto-detect from env vars; switch model per-call.

Auth, RBAC, SSO, audit — in OSS

API key + JWT + OIDC (Google/MS/Okta/GitHub/Keycloak) + 4 default roles + append-only audit log. Most competitors paywall this.

User-delegated OAuth

Each Praxia user authorizes Box / SharePoint / Dropbox / Drive / Salesforce with their own credentials. The external system's native ACL is enforced per Praxia user — alice can only see what alice has access to.

KMS-backed token encryption

OAuth tokens use envelope encryption — fresh DEK per write, AES-GCM payload, DEK wrapped by your KMS. 5 adapters: local / aws / azure / gcp / vault. Master key never lives on the application host.

Production OAuth callback (HTTP)

praxia serve exposes /api/v1/oauth/{provider}/{start,callback,status}. Multi-worker safe state cache (TTL-pruned JSON), pinned redirect URI via PRAXIA_PUBLIC_URL, optional success-redirect to your frontend.

Resource access policies (ACL)

Glob-pattern allow / deny rules per resource type (connector, memory, prompt, skill). Built for enterprise IS departments. Every decision audit-logged.

Autonomous agent — LLM-driven tool-use loop

An LLM-driven tool-use loop over your full Praxia stack — personal memory, org memory, frozen layer, skills, connectors. The agent picks tools on its own (search → run skill → pull connector → answer) with ACL gates and audit logging. Ships as praxia.agent.AutonomousAgent, praxia agent run, and an MCP meta-tool for remote clients.

Prompt Designer — turn intent into a polished template

Describe the task in one line ("score contract risk 1-5 in JSON") → get a production-grade prompt design back: tuned system message, ${variable} user template, 2-3 few-shot examples, 5-criterion rubric. Per-LLM idioms applied automatically (Claude XML / OpenAI JSON-mode / DeepSeek-R1 reasoning / Mistral concise / Llama numbered steps).

Document Designer — code-gen pptx / docx (Claude-Skills-style)

The LLM authors python-pptx / python-docx code, a sandbox runs it (AST allowlist + 30s timeout + 512MB cap on POSIX), and you get a design-rich .pptx / .docx back — multi-column layouts, matrix slides, embedded matplotlib charts, themed branding (colors / fonts / logo / footer from .praxia/themes/). On traceback the error is fed back to the LLM and the attempt repeats up to 3 times. Themes managed in Admin → 🎨 Themes.

Workflow-specialized flows

Sales prep, logic checking, RAG self-correction — three production-ready multi-agent pipelines that run in 5 minutes. No bespoke orchestration code required.

6 default business skills

Investment, sales, design, purchasing, patent, legal — domain-tuned agents with built-in guardrails (tax law, jurisdictional caveats, hallucination guards).

MCP / Claude Skills compatible

Skills serialize to standard SKILL.md. Drop into Claude Skills, Cursor Skills, or any MCP-compatible registry without code changes.

Evidence by default

Sentence-level hallucination detection and retrieval metrics ship as first-class modules. "It works" comes with proof attached.

LLM output quality eval (CI gate)

Catch quality regressions before merge. tests/llm_eval/ grades real LLM output against rubrics + a committed baseline. Score drop > 5pt fails the build. Per-skill cases ship for all 6 skills.

A/B experiments built in

Test prompt variants on real users with deterministic per-user assignment (SHA-256 bucket). Audience filter (roles / users / window). Outcome rollup + tentative winner detection. CLI + SDK.

Hermetic test harness — stubs & drivers for every surface

Every public surface (auth / memory / fusion / exporters / OAuth / parsers / CLI / extensions / experiments / connectors / agent) ships with backend stubs, fixture factories, and protocol-conforming drivers — so contributors can write hermetic tests without standing up real services. CI runs them on every PR.

20 storage / SaaS connectors

Box / SharePoint / Dropbox / Drive / kintone / Salesforce + Notion / Confluence / Jira / Slack / Teams / GitHub / HubSpot / Zendesk / Linear / S3 / Azure Blob / GCS / WebDAV / Email. Per-user OAuth means alice only sees what alice can in each system.

File parsers (PDF · Word · Excel · PowerPoint · CSV · HTML)

Drop a file in — auto-dispatch by extension. PDF page-by-page, Word with heading detection, Excel as Markdown tables, PowerPoint with speaker notes. Custom formats register via entry points.

Output exporters (HTML · PPTX · DOCX · MD · JSON)

Skills produce Markdown by default. OutputFormatSkill infers requested format from natural-language hints ("パワポで" → PPTX, "as a Word doc" → DOCX). Custom formats register via entry-point.

Voice input + voice output

Speech-to-text (Whisper) and text-to-speech (OpenAI TTS / ElevenLabs / Piper). Embedded in Streamlit UI as record-and-go input and read-aloud output.

Full admin user CRUD

Create / update / delete / deactivate / rotate keys / change roles — all via CLI, UI, or SDK. All operations audited.

Admin-controlled LTM policy

Pin which backend(s) users may pick and what the default mode is, at the tenant level. Resolution: admin enforced > call-site > user pref > admin default.

Admin data exports

CSV / JSON / JSONL exports of audit log, users, usage, memory, policies — for compliance, SIEM, backups. Each export action self-audited.

Personal & org dashboards

Flow / skill counts, success rate, top users, promoted blocks, frozen files, distributed skills — out of the box, with no separate analytics pipeline.

Custom prompt distribution

Users save personal prompts. Admins promote them to org or push to specific roles / users. Three scopes with merge precedence.

MCP server (stdio + remote HTTP/SSE)

Use Praxia from Claude Desktop / Cursor / Continue.dev. Local: praxia mcp serve. Remote (multi-host): praxia serve exposes /api/v1/mcp with auth + audit log. Every skill + flow becomes an MCP tool automatically.

Backend-only or full-stack

Use Praxia as a brain behind your own frontend (SDK embed or praxia serve FastAPI HTTP API), or run the bundled Streamlit UI for the fastest path. Same auth, memory, skills.

Apache 2.0 + Open Core ready

Permissive license, commercial-friendly. NOTICE.md inventories every dependency's license. Open Core path for enterprise extras planned.

Mobile-responsive UI + landing

Landing has chip-style nav on phones, scrollable tabs, ≥44px touch targets, prefers-reduced-motion respected. Streamlit UI injects responsive CSS + a "Compact mode" toggle for slow connections.

Sales Agent Flow

Customer IR + past minutes + RAG → hypotheses → FAQ → proposal outline.

praxia run sales \
  --customer-name "Acme" \
  --product "BizFlow"

Logic Checker Flow

Three agents (structure / contradiction / reader) review long docs.

praxia run logic \
  --document spec.md

RAG Optimization Flow

Self-correcting RAG: query expansion → eval → hallucination check loop.

praxia run rag \
  --question "What license?"

Custom flow (~30 lines)

Define a multi-agent pipeline by subclassing Flow. Each step references prior outputs via ${var} templates.

class IncidentResponseFlow(Flow):
    name = "incident_response"
    steps = [
        FlowStep("triage", ...),
        FlowStep("hypothesis", ...),
        FlowStep("mitigation", ...),
    ]

Custom skill (~20 lines)

Subclass Skill with a system prompt + manifest. Auto-serializes to SKILL.md for MCP / Claude Skills.

class HRRecruitingSkill(Skill):
    manifest = SkillManifest(
        name="hr_recruiting",
        domain="hr",
        ...
    )
    system_prompt = """..."""

Custom LTM backend

Implement the 4-method MemoryBackend protocol. Plug in any vector DB (Pinecone, Weaviate, Qdrant, ...) — and optionally combine with built-ins via CompositeBackend / RoutedBackend.

class PineconeBackend:
    def add(...): ...
    def search(...): ...
    def all(...): ...
    def clear(...): ...

Custom connector (~50 lines)

Implement the 2-method Connector protocol (pull / push). Per-user OAuth, ACL enforcement, and audit logging plug in for free. End-to-end Notion example in the guide.

class NotionConnector:
    name = "notion"
    def pull(self, path, *, limit): ...
    def push(self, path, data): ...

Custom output format

Built-in: HTML, PPTX, DOCX, MD, JSON. Add your own (LaTeX? RTF? Confluence Storage?) by implementing the Exporter protocol and declaring an entry-point.

class LatexExporter:
    format = "latex"
    extensions = ("tex",)
    def export(self, content) -> bytes: ...