{"id":50645419,"url":"https://github.com/praxia-dev/praxia","last_synced_at":"2026-06-07T12:02:04.388Z","repository":{"id":356235669,"uuid":"1231466121","full_name":"praxia-dev/praxia","owner":"praxia-dev","description":"Multi-agent orchestrator with cyclic personal-to-org memory · Apache 2.0 · Python 3.11+","archived":false,"fork":false,"pushed_at":"2026-06-05T01:42:46.000Z","size":8204,"stargazers_count":7,"open_issues_count":0,"forks_count":0,"subscribers_count":6,"default_branch":"main","last_synced_at":"2026-06-05T03:11:37.565Z","etag":null,"topics":["ai-agents","apache-2-0","audit-log","autonomous-agents","knowledge-management","llm","mcp","mem0","multi-agent","oauth","python","rag","rbac"],"latest_commit_sha":null,"homepage":"https://praxia.tools","language":"Python","has_issues":true,"has_wiki":null,"has_pages":null,"mirror_url":null,"source_name":null,"license":"apache-2.0","status":null,"scm":"git","pull_requests_enabled":true,"icon_url":"https://github.com/praxia-dev.png","metadata":{"files":{"readme":"README.md","changelog":null,"contributing":"CONTRIBUTING.md","funding":null,"license":"LICENSE","code_of_conduct":"CODE_OF_CONDUCT.md","threat_model":null,"audit":null,"citation":null,"codeowners":".github/CODEOWNERS","security":"SECURITY.md","support":null,"governance":null,"roadmap":null,"authors":null,"dei":null,"publiccode":null,"codemeta":null,"zenodo":null,"notice":"NOTICE.md","maintainers":null,"copyright":null,"agents":null,"dco":null,"cla":null}},"created_at":"2026-05-07T01:46:27.000Z","updated_at":"2026-06-05T01:42:50.000Z","dependencies_parsed_at":null,"dependency_job_id":null,"html_url":"https://github.com/praxia-dev/praxia","commit_stats":null,"previous_names":["praxia-dev/praxia"],"tags_count":0,"template":false,"template_full_name":null,"purl":"pkg:github/praxia-dev/praxia","repository_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/praxia-dev%2Fpraxia","tags_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/praxia-dev%2Fpraxia/tags","releases_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/praxia-dev%2Fpraxia/releases","manifests_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/praxia-dev%2Fpraxia/manifests","owner_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/owners/praxia-dev","download_url":"https://codeload.github.com/praxia-dev/praxia/tar.gz/refs/heads/main","sbom_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/praxia-dev%2Fpraxia/sbom","scorecard":null,"host":{"name":"GitHub","url":"https://github.com","kind":"github","repositories_count":286080680,"owners_count":34020187,"icon_url":"https://github.com/github.png","version":null,"created_at":"2022-05-30T11:31:42.601Z","updated_at":"2026-05-26T15:22:16.424Z","status":"online","status_checked_at":"2026-06-07T02:00:07.652Z","response_time":124,"last_error":null,"robots_txt_status":"success","robots_txt_updated_at":"2025-07-24T06:49:26.215Z","robots_txt_url":"https://github.com/robots.txt","online":true,"can_crawl_api":true,"host_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub","repositories_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories","repository_names_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repository_names","owners_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/owners"}},"keywords":["ai-agents","apache-2-0","audit-log","autonomous-agents","knowledge-management","llm","mcp","mem0","multi-agent","oauth","python","rag","rbac"],"created_at":"2026-06-07T12:02:03.657Z","updated_at":"2026-06-07T12:02:04.365Z","avatar_url":"https://github.com/praxia-dev.png","language":"Python","funding_links":[],"categories":[],"sub_categories":[],"readme":"# Praxia\n[![PyPI](https://img.shields.io/pypi/v/praxia.svg)](https://pypi.org/project/praxia/)\n[![PyPI Downloads](https://img.shields.io/pypi/dm/praxia.svg)](https://pypi.org/project/praxia/)\n[![tests](https://github.com/praxia-dev/praxia/actions/workflows/test.yml/badge.svg)](https://github.com/praxia-dev/praxia/actions/workflows/test.yml)\n[![Follow on X](https://img.shields.io/twitter/follow/praxia_dev?style=social)](https://x.com/praxia_dev)\n\n🌐 **Live**: [praxia.tools](https://praxia.tools/) (primary, Cloudflare) · [praxia-dev.github.io/praxia](https://praxia-dev.github.io/praxia/) (mirror, GitHub Pages) · [@praxia_dev](https://x.com/praxia_dev) on X\n\n![Praxia hero](docs/images/hero-banner.svg)\n\n[![Praxia — 60-second walkthrough](docs/images/demo-thumb.png)](https://youtu.be/o_6NbjJU1AA \"▶ Click to watch (60s)\")\n\n\u003csub\u003e📺 [Watch the 60-second walkthrough](https://youtu.be/o_6NbjJU1AA) · 🚀 [Quickstart](docs/quickstart.md) · 💬 [Discussions](https://github.com/praxia-dev/praxia/discussions)\u003c/sub\u003e\n\n\u003e **Specialized Multi-Agent Orchestrator with Cyclic Personal/Organizational Memory**\n\u003e\n\u003e A workflow-specific multi-agent orchestrator that **automatically promotes** individual tacit knowledge into organizational know-how. Built on a 5-layer memory stack with three independent promotion paths.\n\n[![License: Apache 2.0](https://img.shields.io/badge/License-Apache%202.0-blue.svg)](LICENSE)\n[![Python: 3.11+](https://img.shields.io/badge/python-3.11%2B-blue.svg)](https://www.python.org)\n[![Status: Alpha](https://img.shields.io/badge/status-alpha-orange.svg)]()\n[![Tests: 431](https://img.shields.io/badge/tests-431%20passing-green.svg)]()\n[![Connectors: 20](https://img.shields.io/badge/connectors-20-blueviolet.svg)]()\n[![Languages: 8](https://img.shields.io/badge/languages-en%20%C2%B7%20ja%20%C2%B7%20zh%20%C2%B7%20ko%20%C2%B7%20es%20%C2%B7%20fr%20%C2%B7%20de%20%C2%B7%20pt-blue.svg)]()\n[![MCP: stdio + HTTP/SSE](https://img.shields.io/badge/MCP-stdio%20%2B%20HTTP%2FSSE-orange.svg)]()\n\n\u003e 🔍 Complete feature reference: [docs/FEATURES.md](docs/FEATURES.md)\n\u003e 📊 Concrete Before/After tables: [docs/use-cases.md](docs/use-cases.md)\n\n---\n\n## ⬇ **Try Praxia in 30 seconds — no Python, no setup**\n\n### 👉 [**📦 Download Praxia Desktop for Windows (.exe, 167 MB)**](https://github.com/praxia-dev/praxia/releases/latest/download/Praxia.Desktop_0.1.0-20_x64-setup.exe)\n\nWindows 10 / 11 x64 alpha · Tauri + embedded Python sidecar · zero\n`pip install`, zero `praxia serve`. Paste an LLM API key (Anthropic /\nOpenAI / Azure / Google / Qwen / HF / Ollama) and you're chatting.\nUnsigned alpha — SmartScreen will warn on first launch, click\n**\"More info\" → \"Run anyway\"**.\n\nOther downloads · [`.msi` for managed deployment](https://github.com/praxia-dev/praxia/releases/latest/download/Praxia.Desktop_0.1.0-20_x64_en-US.msi)\n· [all releases \u0026 notes](https://github.com/praxia-dev/praxia/releases/latest)\n· macOS / Linux coming in Phase 1b.\n\n---\n\n## 🎯 Why Praxia?\n\nGeneral-purpose multi-agent frameworks (CrewAI, AutoGen, LangGraph, …) are powerful but stop short on these four problems:\n\n| Problem with existing frameworks | Praxia's approach |\n|---|---|\n| Setup is complex; production deployment is hard | **Workflow-specific templates** (sales prep / logic check / RAG optimization) that run in 5 minutes |\n| Senior-engineer \"magic prompts\" stay locked in one person's editor | **Personal-to-org auto-promotion pipeline** built in |\n| \"It works\" doesn't prove \"it works *well*\" | **Hallucination detection + retrieval evals** shipped by default |\n| Agents stagnate after launch | **Sleep-time consolidation** distills your past flows nightly |\n\nPraxia turns \"one expert's drawer\" into \"everyone's best practices.\"\n\n---\n\n## 👥 Who Praxia is for\n\n| Persona | What they need | How Praxia fits | Typical year-1 result |\n|---|---|---|---|\n| **🏢 Information Systems / Platform team** (300–5,000 employees) | Roll out AI tools without paywalled SSO/RBAC/audit, on-prem option | Auth + RBAC + ACL + per-user OAuth + audit log all in OSS, self-hostable | 100 KW × ~$1.25M net benefit, full audit trail |\n| **🏗️ Engineering / Product VP** (50–500 in scope) | Senior architect bottleneck; junior PM 12–18mo ramp | DesignSkill + sleep-time consolidation distills senior review patterns; Markdown+git frozen layer fits PR workflow | Senior load 16h/wk → 4h/wk; junior ramp 6–9mo |\n| **⚖️ Legal / Compliance lead** (regulated industry) | 50–100 contracts/mo bottleneck; need *auditable* AI workflow without lock-in | LegalSkill (RACE) + read-only memory mode + per-user OAuth + every action audited; Apache 2.0 source for auditors | 60–90min → 10–15min/contract; throughput 50–80/mo → 200–300/mo |\n| **🧪 OSS / Research integrator** | Build domain agent system without re-implementing auth, memory cycling, exporters | 7 plugin types (~50 LoC each); use as library, run `praxia serve` as backend, embed in LangGraph | Day-30: domain skill + custom connector + memory cycling working — **~3 weeks ahead of from-scratch** |\n\nDetailed Before/After by industry: [docs/use-cases.md](docs/use-cases.md).\n\n---\n\n## 💎 Why OSS matters here\n\nThe capabilities you typically pay enterprise tier for — already in the Apache 2.0 package:\n\n- **SSO + RBAC + audit are not paywalled.** OIDC SSO (Google / Microsoft / Okta / GitHub / Keycloak) is in the OSS. Most agent frameworks ship without it; most agent platforms paywall it. Praxia treats it as table stakes.\n- **Memory format is not locked in.** Layer 4 is plain Markdown in your git repo. Layer 3 exports to JSONL. Layer 1 is your chosen backend's native format. Leaving costs nothing.\n- **You can read every line.** Apache 2.0. Show the source to your auditors, your security team, your customers.\n- **Multi-LTM ensembles, not single-vendor.** Run Mem0 + Zep + HindSight in parallel, fuse with RRF, or route per query. No commercial agent platform exposes this — they pick a backend and lock you in.\n- **Per-user OAuth respects external ACL.** When Alice pulls from Box, Box's own ACL applies — Alice only sees what Alice can see. Service-account designs (typical SaaS shortcut) leak data across users.\n- **Air-gapped operation.** `PRAXIA_LOCAL_MODEL=gemma`, Ollama, `backend=json` — no cloud LLM, no cloud vector DB, no telemetry. Same code as cloud customers.\n- **Production-grade OAuth + KMS in OSS.** Multi-worker safe state cache, 5 KMS adapters (AWS / Azure / GCP / Vault / local). Most agent platforms paywall this; Praxia ships it.\n- **A/B experiments + quality eval included.** Test prompt variants on real users with deterministic assignment; catch LLM output quality regressions in CI.\n\n---\n\n## 🏗 Architecture — 5-Layer Memory Stack\n\n![Architecture diagram](docs/images/architecture.svg)\n\nThe same picture as ASCII art:\n\n```\n┌──────────────────────────────────────────────────────────┐\n│  AI Agents (Skills + MCP)                                │\n└──────────────┬───────────────────────────────────────────┘\n               │ Users just have normal conversations\n               ▼\n╔═══════════════════════════════════════════════════════════╗\n║ Layer 1: Personal memory (auto-extracted)                 ║\n║   Mem0 / LangMem / HindSight / Letta / Zep / JSON         ║\n║   namespace = user_id                                     ║\n║   ★ Zero-effort tacit-knowledge capture                   ║\n╚══════════════╤════════════════════════════════════════════╝\n               │ Sleep-time Consolidation (nightly batch)\n               ▼\n╔═══════════════════════════════════════════════════════════╗\n║ Layer 2: Distillation \u0026 promotion engine                  ║\n║   Three parallel \"validity tests\":                        ║\n║     ① Frequency  (recurring across N+ users)              ║\n║     ② Outcome    (correlated with wins/losses)            ║\n║     ③ Self-eval  (LLM scored)                             ║\n╚══════════════╤════════════════════════════════════════════╝\n               │ Auto-promote above threshold; queue otherwise\n               ▼\n╔═══════════════════════════════════════════════════════════╗\n║ Layer 3: Shared memory (living organizational knowledge)  ║\n║   Letta-style shared blocks; all agents read/write        ║\n╚══════════════╤════════════════════════════════════════════╝\n               │ PR review for high-impact items\n               ▼\n╔═══════════════════════════════════════════════════════════╗\n║ Layer 4: Frozen layer (git-managed best practices)        ║\n║   Markdown + git + PR review                              ║\n║   GitHub Copilot / Cursor Rules-compatible format         ║\n╚══════════════╤════════════════════════════════════════════╝\n               │ (optional)\n               ▼\n╔═══════════════════════════════════════════════════════════╗\n║ Layer 5: Graph layer (only relationship-heavy domains)    ║\n║   Zep / Graphiti — decisions, customer 360, incident DAG  ║\n╚═══════════════════════════════════════════════════════════╝\n\nParallel Layer 6: Skills registry\n  Personal skills get promoted to the organizational catalog.\n  MCP / Claude Skills / Cursor Skills compatible.\n```\n\nThree promotion paths (**auto / statistical / manual**) run side by side — never depending on a single mechanism.\n\nFor details, see [docs/architecture.md](docs/architecture.md).\n\n---\n\n## ✨ What's Bundled\n\n### Autonomous agent (LLM-driven tool-use loop)\n\n`praxia.agent.AutonomousAgent` runs an LLM-driven tool-use loop over the\nfull Praxia stack — personal/org memory, skills, frozen layer, connectors —\nwith ACL checks and audit logging built in. The LLM picks tools on its own\nuntil it has the information it needs, mirroring how modern code-editing assistants drives its\nown tool use.\n\n```python\nfrom praxia.agent import AutonomousAgent\nfrom praxia.core.llm import LLM\n\nagent = AutonomousAgent(user_id=\"alice\", org_id=\"acme\", llm=LLM(\"claude\"))\nresult = agent.run(\"Tell me what we know about Acme and draft a proposal.\")\nprint(result.final_text)\n```\n\n```bash\npraxia agent run \"Summarize where we stand with Acme this quarter and draft a proposal\"\npraxia agent tools     # list the 11 built-in tools\n```\n\nThe agent is also exposed as a single MCP meta-tool (`autonomous_agent`) so\nremote clients (Claude Desktop, Cursor) can delegate an entire investigation\nwithout orchestrating individual tools by hand. See\n[FEATURES § 38](docs/FEATURES.md#38-autonomous-agent-llm-driven-tool-use-loop).\n\n### CommandedAgent — autonomous agent with external verification\n\n`AutonomousAgent` is a free-running tool-use loop — perfect when the\nenvironment *is* the answer key (tests pass / fail, commands exit 0 /\nnon-zero). For workloads where the environment doesn't give you that\nfree check — private-corpus fact QA, SOP / compliance, customer\nsupport over manuals, technical-knowledge transfer — `CommandedAgent`\nwraps it with three guards: **pre-retrieval + grounding verification +\nbounded retry**, with an explicit `abstain` path when the sources\ndon't support a confident answer.\n\nCalibrated against an in-house multi-hop RAG harness\n(HotpotQA / SQuAD v2 / JEMHopQA) — see\n[`docs/VERIFICATION_FINDINGS.md`](docs/VERIFICATION_FINDINGS.md) — and\nthe resulting defaults:\n\n- **Task-type router** — `default_task_classifier` sends coding /\n  command / tool prompts straight through (the environment is the\n  verifier for those), and routes knowledge-QA through the grounding\n  gate. Bilingual (EN/JA) keyword classifier; pluggable.\n- **Query decomposition** — wire a `QueryDecomposer` into the\n  retriever and multi-hop questions are split into sub-questions,\n  retrieved per hop, and unioned (deduped). +12pt on HotpotQA-distractor\n  40q in the source benchmark; single-hop queries pass through\n  unchanged.\n- **No-improvement early stop** — a redraft that fails to lift\n  groundedness by `min_groundedness_improvement` (default 0.05) aborts\n  to abstain instead of burning the whole round budget on a loop that\n  isn't making progress.\n- **Promotion engine reweighted** — `PromotionEngine` defaults are now\n  `0.5·freq + 0.4·outcome + 0.1·self_eval` so the LLM's own\n  self-assessment can no longer carry a memory promotion on its own.\n\n```python\nfrom praxia.agent import AutonomousAgent, CommandedAgent\nfrom praxia.agent.decomposer import LLMQueryDecomposer\nfrom praxia.agent.commander import DefaultMemoryRetriever\nfrom praxia.core.llm import LLM\n\nllm = LLM(\"claude\")\ninner = AutonomousAgent(user_id=\"alice\", org_id=\"acme\", llm=llm)\nretriever = DefaultMemoryRetriever(\n    personal=inner.memory,                  # L1\n    decomposer=LLMQueryDecomposer(llm=llm), # K2: multi-hop split\n)\nagent = CommandedAgent(inner, retriever=retriever, max_verify_rounds=3)\n\nresult = agent.run(\n    \"How do we handle Customer X's stamping-press alarm code E-204?\"\n)\n# result.answer carries [L1#0, L3#2, ...] citations\n# result.verdict.decision ∈ {\"accept\", \"redraft\", \"abstain\"}\n# result.stopped_reason ∈ {\"accept\", \"abstain\", \"no_improvement\",\n#                          \"max_rounds\", \"bypass_action\"}\n# result.task_kind         ∈ {\"knowledge\", \"action\"}\n# result.rounds            is the per-round draft + verdict trace\n```\n\nPluggable everything: `Verifier` / `Retriever` / `QueryDecomposer` /\n`TaskClassifier` are all protocols or callables, so you can drop in\nTiDB Vector / pgvector / hybrid BM25+ANN / GraphRAG or your own\ngrounding scorer / decomposer without touching core. See\n[FEATURES § 38b](docs/FEATURES.md#38b-commandedagent--autonomous-agent-with-external-grounding-commander)\nand [`docs/COMMANDED_AGENT.md`](docs/COMMANDED_AGENT.md).\n\n### 3 Specialized Multi-Agent Flows\n\n| Flow | What it does |\n|---|---|\n| **SalesAgentFlow** | Reads customer IR, past minutes, RAG context → generates **hypotheses → FAQ → proposal outline** |\n| **LogicCheckerFlow** | Three agents (structure / contradiction / reader) review long documents for logical consistency |\n| **RAGOptimizationFlow** | Self-correcting RAG: query expansion → retrieval → relevance eval → hallucination check loop |\n\n### 6 Default Business-Domain Skills\n\n| Skill | Domain | Use cases |\n|---|---|---|\n| **InvestmentSkill** | Investment | Equity research, due diligence, portfolio decisions |\n| **SalesSkill** | Sales | Account research, proposal drafting, FAQ prep |\n| **DesignSkill** | Engineering Design | System design review, requirements engineering |\n| **PurchasingSkill** | Procurement | Supplier evaluation, RFQ analysis, TCO, BCP risk |\n| **PatentSkill** | IP / Patent | Prior-art search, claims drafting, patent maps |\n| **LegalSkill** | Legal | Contract review, compliance, M\u0026A diligence |\n\nEach skill serializes to Claude-Skills / MCP-compatible `SKILL.md`.\n\nPlus four **utility** skills:\n\n| Skill | What it does |\n|---|---|\n| **`PromptDesignerSkill`** | Take a one-line task description → produce a production-grade prompt template (system + user + 2-3 few-shot examples + 5-criterion rubric) tuned for the target LLM (Claude / OpenAI / DeepSeek / Mistral / Llama / …). Save to `PromptStore`, A/B-test via `praxia.experiments`. |\n| **`OutputFormatSkill`** | Detect \"export as PowerPoint\" / \"as Word doc\" / etc. in natural language (English + JA / ZH / KO / ES / FR / DE / PT-BR phrases also recognized) and dispatch to the matching exporter (PPTX / DOCX / HTML / MD / JSON). |\n| **`PptxDesignerSkill`** | **Code-gen** path (Claude-Skills-style): the LLM authors `python-pptx` code that runs in a sandbox to produce a design-rich `.pptx` — multi-column layouts, matrix slides, embedded charts, themed branding. Themes (colors / fonts / logo / footer) live under `.praxia/themes/\u003cname\u003e/` and are managed in `Admin → 🎨 Themes`. |\n| **`DocxDesignerSkill`** | Same approach for Word documents — LLM-authored `python-docx` code, sandbox execution, themed `.docx` output (heading hierarchy, page footer, tables, callouts, embedded charts). |\n\n```bash\n# Generate a prompt template for any task\npraxia skill run prompt_designer \"Have in-house legal score contract risk on a 5-point scale\"\n\n# Generate a design-rich slide deck via code-gen\npraxia skill run pptx_designer \"Q4 sales review for Acme — cover, exec summary, top 3 customers, 2x2 challenges matrix, next actions. 10 slides.\"\n```\n\n### All Major LLMs\n\nLiteLLM-powered single-line provider switching:\n\n| Provider | Aliases (current default) | Auth env var(s) |\n|---|---|---|\n| Anthropic Claude | `claude` (Opus 4.7) · `claude-sonnet` (4.6) · `claude-haiku` (4.5) | `ANTHROPIC_API_KEY` |\n| OpenAI ChatGPT | `chatgpt` (GPT-5.5) · `gpt-5.5` · `gpt-5.4` · `gpt-5` · `o4-mini` · `o3` · `gpt-4o` | `OPENAI_API_KEY` |\n| **Azure OpenAI Service** | `azure/\u003cdeployment-name\u003e` | `AZURE_API_KEY` + `AZURE_API_BASE` + `AZURE_API_VERSION` (auto-mirrored to AZURE_OPENAI_*) |\n| **Azure AI Foundry (Inference)** | `azure_ai/\u003cmodel-name\u003e` | `AZURE_AI_API_KEY` + `AZURE_AI_API_BASE` |\n| **AWS Bedrock** | `bedrock/anthropic.claude-opus-4-7-v1:0` etc. | `AWS_ACCESS_KEY_ID` + `AWS_SECRET_ACCESS_KEY` + `AWS_REGION` |\n| **Google Vertex AI** | `vertex_ai/gemini-3.1-pro` · `vertex_ai/claude-opus-4-7` | `VERTEX_PROJECT` + `VERTEX_LOCATION` + `GOOGLE_APPLICATION_CREDENTIALS` |\n| Google Gemini (public API) | `gemini` (3.1 Pro) · `gemini-flash-lite` (3.1) · `gemini-2.5-pro` | `GEMINI_API_KEY` |\n| Google Gemma (open) | `gemma` / `gemma-2b` / `gemma-9b` / `gemma-27b` (Ollama) · `gemma-cloud` (Vertex AI) | (none for local) / Vertex auth |\n| Alibaba Qwen (cloud) | `qwen` / `qwen-72b` | `DASHSCOPE_API_KEY` |\n| **DeepSeek** | `deepseek` (V4) · `deepseek-v4-pro` · `deepseek-v3.2` · `deepseek-reasoner` (V3.2 Speciale) | `DEEPSEEK_API_KEY` |\n| **Mistral** | `mistral` (Large 3) · `mistral-medium-3.5` · `mistral-small-4` · `magistral` (reasoning) · `codestral` (25.08) · `devstral` (agentic code) | `MISTRAL_API_KEY` |\n| **xAI Grok** | `grok` (Grok 4.1) · `grok-4.1-fast` · `grok-4-heavy` · `grok-code` (code-tuned) | `XAI_API_KEY` |\n| **Llama (Groq fast)** | `llama` (3.3 70B Versatile via Groq) | `GROQ_API_KEY` |\n| **Cohere** | `command-r` (Command R+) | `COHERE_API_KEY` |\n| **Perplexity Sonar** | `perplexity` (Sonar Pro, web search) · `perplexity-cheap` · `perplexity-reasoning` | `PERPLEXITY_API_KEY` |\n| **Microsoft Phi** (local) | `phi` (3.5 3.8B Ollama) | (none — runs in-house) |\n| Qwen / Llama / Phi (local) | `qwen-local` · `llama-local` · `phi` (Ollama) | (none — runs in-house) |\n\n```python\nLLM(\"claude\")          # Anthropic Claude Opus 4.7\nLLM(\"chatgpt\")         # OpenAI GPT-5.5\nLLM(\"gemini\")          # Gemini 3.1 Pro\nLLM(\"grok\")            # xAI Grok 4.1\nLLM(\"deepseek\")        # DeepSeek V4 — strong + low cost\nLLM(\"mistral\")         # Mistral Large 3 — EU-friendly\nLLM(\"perplexity\")      # Sonar Pro — web-search-augmented\nLLM(\"azure/gpt-5.5\")   # Azure OpenAI deployment named \"gpt-5.5\"\nLLM(\"bedrock/anthropic.claude-opus-4-7-v1:0\")  # Claude on AWS Bedrock\nLLM(\"vertex_ai/gemini-3.1-pro\")                # Gemini on Vertex\nLLM(\"openai/gpt-5.5\")  # Any LiteLLM-compatible model string\n```\n\nThe Admin → Settings UI groups all these as provider sub-categories (`LLM · OpenAI`, `LLM · Azure OpenAI`, etc.), with required-key indicators and per-key delete. The provider picker has a `Custom deployment / model name…` entry per-provider so deployment-named Azure/Bedrock/Vertex models work without leaving the Provider context. `litellm.drop_params + modify_params` are enabled at import time so per-model quirks (GPT-5 rejects `temperature`, requires `max_completion_tokens` instead of `max_tokens`, etc.) are bridged transparently.\n\n### File parsing — PDF · Office · CSV · TXT · HTML · MD · code · vision-enriched\n\n\u003e **alpha20+**: PDF / DOCX / PPTX parsers also surface embedded /\n\u003e rendered images so the vision LLM can read charts, diagrams,\n\u003e screenshots, and other figures inside the documents — not just the\n\u003e surrounding text. PDF pages get rasterized to JPEG at 100 DPI;\n\u003e DOCX / PPTX have their embedded media extracted from the ZIP's\n\u003e `word/media/` and `ppt/media/` folders. The Desktop chat composer\n\u003e accepts these formats inline too: drop a PDF, ask \"what does the\n\u003e chart on page 3 say?\", and Praxia routes the page image to the\n\u003e vision LLM automatically.\n\n\nAuto-dispatched by extension:\n\n| Extension | Parser | Optional dep |\n|---|---|---|\n| `.txt` `.md` `.rst` `.py` `.ts` `.js` | TextParser | (none) |\n| `.csv` `.tsv` | CsvParser | (stdlib) |\n| `.json` `.yaml` `.yml` | StructuredParser | (core) |\n| `.html` `.xml` | HtmlParser | (stdlib) |\n| `.pdf` | PdfParser | `praxia[office]` |\n| `.docx` | DocxParser | `praxia[office]` |\n| `.pptx` | PptxParser | `praxia[office]` |\n| `.xlsx` `.xlsm` | XlsxParser | `praxia[office]` |\n\n```python\nfrom praxia.io.parsers import parse_file\n\ndoc = parse_file(\"contract.pdf\")          # works\ndoc = parse_file(\"Q3_results.xlsx\")       # also works\nprint(doc.content)\n```\n\nThird-party formats register via `[project.entry-points.\"praxia.parsers\"]` — no fork required.\n\n### Output exporters — render skill output to HTML / PPTX / DOCX / MD / JSON\n\nSkills produce Markdown by default. Convert to whatever the user requested:\n\n```python\nfrom praxia.io.exporters import export_as\nresult = export_as(md_text, format=\"pptx\", title=\"Q3 Review\")\n# result.bytes → write to disk, stream over HTTP, push via a connector\n```\n\n`OutputFormatSkill` infers the format from natural-language hints across multiple languages:\n\n```python\nfrom praxia.skills.output_format import OutputFormatSkill\nfs = OutputFormatSkill()\nfs.detect(\"export as PowerPoint\").format      # → \"pptx\"\nfs.detect(\"as a Word document\").format        # → \"docx\"\nfs.deliver(md, user_request=\"HTML please\")    # ExporterResult with .bytes\n```\n\nCLI shortcut:\n```bash\npraxia export report.md report.html\npraxia export report.md slides.pptx --title \"Q3 Review\"\n```\n\nCustom formats register via the `praxia.exporters` entry-point — same pattern as connectors.\n\n### Memory mode — accumulate or read-only, per user\n\nSome sessions shouldn't leave a trail (legal review, sensitive data exploration). Toggle per-user:\n\n```bash\npraxia memory mode --user-id alice read_only      # writes silently dropped\npraxia memory mode --user-id alice accumulate     # back on\npraxia memory show --user-id alice                # see the resolved config + reason\n```\n\nAdmins can lock the mode for the whole tenant or for specific roles:\n```bash\npraxia admin memory-policy-set --default-mode read_only --mode-locked\npraxia admin memory-policy-set --enforced-backend mem0 --allowed mem0,zep\npraxia admin memory-policy-set --accumulate-locked-roles operator,admin\n```\n\nResolution order: admin enforced \u003e call-site argument \u003e user pref \u003e admin default. See [`praxia.memory.policy`](praxia/memory/policy.py).\n\n### Multi-LTM fusion + dynamic routing (accuracy boost)\n\nEach LTM has different strengths — entity linking (Mem0), temporal KG (Zep), audit trail (JSON), vector recall (HindSight). You can run several at once and either fuse the results or pick per-query:\n\n```python\nfrom praxia.memory.composite import CompositeBackend, WeightedBackend\nfrom praxia.memory.router import RoutedBackend, RuleRouter\n\n# A. Parallel fan-out + Reciprocal Rank Fusion\ncomposite = CompositeBackend(\n    backends=[WeightedBackend(\"mem0\", ..., weight=1.5),\n              WeightedBackend(\"zep\", ..., weight=1.0),\n              WeightedBackend(\"hindsight\", ..., weight=1.0)],\n    fusion=\"rrf\",\n)\n\n# B. Query-aware dispatch (RuleRouter handles English + Japanese keywords)\nrouted = RoutedBackend(\n    backends={\"mem0\": ..., \"zep\": ..., \"hindsight\": ..., \"json\": ...},\n    router=RuleRouter(),\n    write_to=\"mem0\",\n)\n```\n\nFull design + tradeoffs: [docs/FEATURES.md § 5.1](docs/FEATURES.md#51-multi-ltm-fusion--dynamic-routing-accuracy-boost).\n\n### Voice input / output\n\n```python\nfrom praxia.io.audio import STT, TTS\n\ntext = STT().transcribe(audio_bytes, filename=\"meeting.wav\", language=\"ja\")\naudio = TTS().synthesize(\"Hello world\", voice=\"alloy\", format=\"mp3\")\n```\n\nBoth Streamlit UI tabs (Run Flow, Skill) include `🎙 Audio input` and `🔊 Read response aloud` toggles. Providers: OpenAI Whisper / TTS (default), ElevenLabs (premium voices), local Whisper / Piper (`praxia[audio-local]`).\n\n### 6 Pluggable LTM Backends\n\n| Backend | Notes |\n|---|---|\n| **json** (default) | Zero-dependency, JSONL on disk, fully auditable |\n| **mem0** | Entity linking + hybrid search (recommended for production) |\n| **langmem** | LangChain LangMem SDK |\n| **letta** | Letta shared blocks (with read-only policy support) |\n| **zep** | Zep / Graphiti for temporal KGs (Layer 5) |\n| **hindsight** | [vectorize-io/hindsight](https://github.com/vectorize-io/hindsight) — agent memory store |\n\nSwitch with one line:\n```python\nPersonalMemory(user_id=\"alice\", backend=\"mem0\")\n```\n\n### Built-in Authentication, RBAC, SSO \u0026 Resource Policies\n\n- **API-key + JWT auth** (`praxia.auth`) with 4 default roles (`admin` / `operator` / `member` / `viewer`)\n- **SSO via OIDC**: Google, Microsoft Entra ID, Okta, GitHub, Keycloak, custom OIDC, plus SAML skeleton\n- **Resource access policies (ACL)** — glob-pattern allow/deny rules per resource (built for enterprise IS departments)\n- **Append-only audit log** — every authn / authz / policy decision / privileged action recorded\n- **Admin data exports** — CSV / JSON / JSONL dumps of audit, users, usage, memory, policies, shared blocks (chain-of-custody preserved)\n\n### Admin User Management\n- Create / read / update / delete users\n- Activate / deactivate, role grants, API-key rotation\n- All actions audited\n- Available via CLI, Streamlit UI, and SDK\n\n### Custom Prompts (per-user + admin-distributed)\n- Users save personal prompts; admins promote them to org or distribute to specific users / roles\n- Three scopes (personal / org / distributed) with merge precedence\n- Same model as the skill registry\n\n### Per-user OAuth for connectors\n\nEach Praxia user can authorize external systems with **their own credentials** — the external system's native ACL is enforced per-user.\n\n```bash\n# Set the OAuth app credentials once\nexport PRAXIA_OAUTH_BOX_CLIENT_ID=...\nexport PRAXIA_OAUTH_BOX_CLIENT_SECRET=...\n\n# Each user authorizes individually (CLI loopback)\npraxia oauth start box --user-id alice\n# → opens authorization URL → user logs in → redirect captures code\n# → token saved encrypted to .praxia/auth/oauth_tokens.jsonl\n\n# From now on, alice's connector calls use her token\npraxia connector pull box 0 --user-id alice\n# alice can only see Box folders alice has access to\n```\n\nSupported providers (13): **Box**, **Microsoft** (SharePoint/OneDrive/Teams), **Dropbox**, **Google Drive**, **Salesforce**, **Notion**, **Atlassian** (Confluence+Jira), **Slack**, **GitHub**, **HubSpot**, **Zendesk**, **Linear**, **kintone**. Tokens auto-refresh; access logged in audit log.\n\n**End-user UI**: each user manages their own connections under **個人設定 → 外部サービス連携** (Preferences → Service Connections) in the Streamlit UI — one row per provider with status (Connected / Not connected / Expired), Connect button (opens IdP login) and Disconnect button (revokes + clears local token).\n\n**Production HTTP callback** (`praxia serve`): four endpoints under `/api/v1/oauth/{provider}/`: `start`, `callback`, `status`, `revoke` (DELETE). State cache is multi-worker-safe (`PersistentStateStore` — TTL-pruned JSON), so the IdP redirect can land on any FastAPI worker. Set `PRAXIA_PUBLIC_URL` to pin the redirect URI.\n\n### KMS-backed token encryption (production)\n\nOAuth tokens use **envelope encryption**: a fresh 256-bit data key per write, AES-GCM payload encryption, and the data key wrapped by a configurable `KmsAdapter`. The master key never lives on the application host:\n\n| Adapter | Install | Use |\n|---|---|---|\n| `local` (default) | (none) | dev / single-host |\n| `aws` | `pip install 'praxia[kms-aws]'` | AWS KMS CMK |\n| `azure` | `pip install 'praxia[kms-azure]'` | Azure Key Vault Keys |\n| `gcp` | `pip install 'praxia[kms-gcp]'` | GCP Cloud KMS |\n| `vault` | `pip install 'praxia[kms-vault]'` | HashiCorp Vault Transit |\n\n```bash\nexport PRAXIA_KMS_ADAPTER=aws\nexport PRAXIA_KMS_KEY_ID=arn:aws:kms:us-east-1:111122223333:key/...\n```\n\nLegacy v0.1 tokens decode transparently — re-saving rewrites in the new envelope format.\n\n### A/B experiments — prompts / skills / LLMs\n\nTest variants of any payload (system prompt, LLM provider, memory backend) with deterministic per-user assignment + outcome tracking:\n\n```bash\npraxia experiment create proposal_v2 \\\n    --name \"Proposal: shorter vs longer prompt\" \\\n    --variants '{\"control\":{\"prompt\":\"\u003c800-word\u003e\"},\"candidate\":{\"prompt\":\"\u003c400-word\u003e\"}}' \\\n    --traffic-split \"control=0.5,candidate=0.5\"\npraxia experiment start proposal_v2\n# ...users run flows; outcomes recorded automatically...\npraxia experiment results proposal_v2\n# → 🏆 Tentative winner: candidate (confidence 0.41)\n```\n\nSame user always sees the same variant during the experiment (SHA-256 bucket). Audience filter (roles / users / time window). See [`praxia.experiments`](praxia/experiments/).\n\n### LLM output quality evaluation\n\nSeparate from the deterministic regression suite — `tests/llm_eval/` runs **real LLM calls** and grades output against rubrics + a committed baseline. CI flags PRs where quality drops \u003e 5 points:\n\n```bash\n# Skipped by default (requires API keys + costs tokens)\npytest tests/llm_eval -m llm_eval -v\n\n# Update baselines after a known-good change\npytest tests/llm_eval --update-baselines\n\n# Compare providers on the same cases\npytest tests/llm_eval --llm-eval-model gpt-4o\n```\n\nBuilt-in rubrics: keyword match, structure (heading) match, length band, must-not-contain, LLM-as-judge. One canonical case per business skill ships out of the box.\n\n### External Connectors — 20 systems, Pull + Push\n\n**Storage / Files (8)**\n\n| Connector | Pull | Push | Auth |\n|---|---|---|---|\n| **Box** | ✅ folder → files | ✅ upload | OAuth2 / JWT |\n| **SharePoint / M365** | ✅ | ✅ | Microsoft Entra |\n| **Dropbox** | ✅ | ✅ | OAuth2 |\n| **Google Drive** | ✅ | ✅ | OAuth / SA |\n| **AWS S3** | ✅ bucket/prefix | ✅ object upload | IAM (boto3 chain) |\n| **Azure Blob Storage** | ✅ | ✅ | DefaultAzureCredential / connstr / SAS |\n| **GCS** | ✅ | ✅ | ADC / service account |\n| **WebDAV / Nextcloud** | ✅ | ✅ | HTTP Basic |\n\n**Knowledge / Docs (3)**\n\n| Connector | Pull | Push | Auth |\n|---|---|---|---|\n| **Notion** | ✅ database query | ✅ child page | OAuth (Notion) |\n| **Confluence** | ✅ CQL search | ✅ child page | OAuth (Atlassian) |\n| **Jira** | ✅ JQL search | ✅ create issue | OAuth (Atlassian) |\n\n**Communication (3)**\n\n| Connector | Pull | Push | Auth |\n|---|---|---|---|\n| **Slack** | ✅ history / search | ✅ post message | OAuth (Slack) |\n| **Microsoft Teams** | ✅ channel messages | ✅ post message | OAuth (Microsoft) |\n| **Email** (IMAP / Gmail / Outlook) | ✅ folder + query | ✅ send | IMAP/SMTP / Google / Microsoft OAuth |\n\n**CRM / Tickets / Engineering (5)**\n\n| Connector | Pull | Push | Auth |\n|---|---|---|---|\n| **kintone** | ✅ | ✅ | OAuth / API token |\n| **Salesforce** | ✅ SOQL | ✅ sObject create | OAuth |\n| **HubSpot CRM** | ✅ contacts/companies/deals | ✅ note attach | OAuth |\n| **Zendesk** | ✅ ticket search | ✅ create ticket | OAuth or API token |\n| **GitHub** | ✅ issues/code/files | ✅ issue / comment | OAuth (GitHub) |\n| **Linear** | ✅ issues by team | ✅ create issue | OAuth or API key |\n\nPull data into agent flows; push agent outputs back to your system of record. All access subject to admin policies. **Per-user OAuth** means alice only sees what alice has access to in each system.\n\n### Dashboards\n- **Personal**: 3 headline KPIs (total runs · success rate · memory entries) + Top skills horizontal bar chart\n- **Organizational**: 3 headline KPIs (active users · org runs · success rate) + Top users / Top skills side-by-side bar charts\n- Charts use plotly with the Praxia gold palette; tabular fallback if plotly isn't installed\n\n---\n\n## 🖼 UI Tour\n\nThe bundled Streamlit UI puts the **non-power-user surface area** of Praxia on a clean 3-zone layout: a sticky top-bar navigation, a sidebar dedicated to data context, and a workspace per view.\n\n```\n┌──────────────────────────────────────────────────────────────────────┐\n│  [🎬 Run] [📝 Prompts] [📁 Data] [🧠 Knowledge] [📊 Dashboard]        │ ← fixed top bar\n│                  [👤 Preferences] [⚙ Admin]*                         │   (* admin role only)\n├────────────────────────┬─────────────────────────────────────────────┤\n│  🪡 Praxia              │                                             │\n│  👤 alice · admin       │                                             │\n│  [Sign out]             │   Selected view's workspace                 │\n│  ───────────────        │                                             │\n│  📁 Context             │   (Run = Agent chat or Skill form,          │\n│  ☑ Personal memory      │    Prompts = designer + library,            │\n│  ☑ Org memory           │    Data = folder CRUD + sharing,            │\n│  ☐ Frozen layer         │    Knowledge = memories + skill registry,   │\n│  📁 Q3 Sales (12)       │    Dashboard = KPIs + charts,               │\n│  🔌 Box: /Customers     │    Preferences = language/theme,            │\n│                         │    Admin = settings/users/policies/...)     │\n├────────────────────────┴─────────────────────────────────────────────┤\n│  [Type a message...                                       📎 📤]      │ ← chat input fixed bottom\n└──────────────────────────────────────────────────────────────────────┘\n```\n\n**Login**:\n- **API key** path — User ID + Password (= API key from `praxia user create`) for role-gated multi-user.\n- **SSO** path — when `PRAXIA_SSO_PROVIDER` is set (Google / Microsoft Entra / Okta / GitHub / Keycloak / generic OIDC), a primary \"Sign in with \\\u003cprovider\\\u003e\" button is rendered above the API-key form; click → IdP redirect → callback restores session.\n- Cookie-based persistent login (30-min sliding TTL, server-side session record at `.praxia/sessions/\u003ctoken\u003e.json`) survives browser reloads — F5 doesn't kick the user back to the login form.\n\n**Run** is the high-frequency view with two sub-tabs:\n- **🤖 Agent** — chat backed by `AutonomousAgent`. Type a goal; the LLM picks tools (search, connectors, skills) and iterates. **Vision attachments** via the 📎 button (PNG/JPG/GIF/WebP) when the model supports it. **Persistent threads**: every conversation saved at `.praxia/chats/\u003cuser\u003e/\u003cid\u003e.json`; resume / rename / delete from the `💬 Conversations` popover. Selected Context folders are passed in as reference data, with grep-relevance filtering on large folders. Layout is ChatGPT-style — top nav fixed top, chat input fixed bottom.\n- **🛠 Skill** — pick a domain skill (investment / sales / design / purchasing / patent / legal), fill the input, click Run. Single-call, single-answer.\n\n**Prompts** has the PromptDesigner (1-line task → polished prompt) and a per-user prompt library with admin-distributed bundles.\n\n**Data folders** are how you create/manage local-upload folders or register external paths (Box / SharePoint / Notion / etc.). Folders can be **shared read-only with other users** (multiselect from Admin → Users — owner can upload + delete + reshare; sharees can only read). Image uploads (PNG/JPG/GIF/WebP) become first-class scope content via the built-in `ImageParser` and feed vision-capable agents.\n\n**Knowledge** shows browseable personal + shared memory, plus the skill registry (your skills + org-promoted ones).\n\n**Dashboard** shows 3 headline KPIs (runs / success-rate / memory entries) and Plotly bar charts of top skills + top users.\n\n**Admin** (admin role only) consolidates 7 sub-tabs: **Settings** (default LLM model + persistent KNOWN_KEYS grouped per provider — `LLM · OpenAI`, `LLM · Anthropic`, `LLM · Azure OpenAI`, `LLM · Azure AI Foundry`, `LLM · AWS Bedrock`, `LLM · Google`, `LLM · DeepSeek`, `LLM · Mistral`, `LLM · xAI`, `LLM · Cohere`, `LLM · Perplexity`, `LLM · Groq`, `LLM · Local (Ollama)`, plus **Memory policy** with `single` / `composite` / `routed` strategy radio — fan reads across multiple backends with RRF fusion, or route per-query via rule/llm router), Users, Connectors, Policies (ACL), Consolidate (sleep-time promotion), Exports (audit / users / memory / policies CSV/JSON/JSONL), About. Set keys are masked as `****` (no partial leak); per-row `🗑 Delete` checkbox removes a key from `.praxia/config.toml`.\n\nCLI users get the same functionality with rich-formatted output:\n\n![CLI terminal](docs/images/cli-terminal.svg)\n\n---\n\n## 🚀 Quickstart\n\n\u003e ### 🖥 Native desktop app (no Python required)\n\u003e\n\u003e The easiest way to try Praxia is the native desktop installer.\n\u003e\n\u003e | Platform | Installer | Status |\n\u003e |---|---|---|\n\u003e | Windows 10 / 11 | `.exe` (NSIS, 140 MB) / `.msi` (WiX, 203 MB) | ✅ **`v0.1.0-alpha1` shipped** |\n\u003e | macOS 12+ | `.dmg` | 🚧 next alpha drop |\n\u003e | Linux (Debian / Ubuntu) | `.deb` / `.AppImage` | 🚧 next alpha drop |\n\u003e\n\u003e 📦 **Direct download (Windows, ~147 MB):**\n\u003e [`Praxia.Desktop_0.1.0-20_x64-setup.exe`](https://github.com/praxia-dev/praxia/releases/latest/download/Praxia.Desktop_0.1.0-20_x64-setup.exe)\n\u003e · alternatives: [`.msi`](https://github.com/praxia-dev/praxia/releases/latest/download/Praxia.Desktop_0.1.0-20_x64_en-US.msi) for managed deployment ·\n\u003e [all releases \u0026 notes](https://github.com/praxia-dev/praxia/releases/latest).\n\u003e The installer is unsigned during alpha, so Windows SmartScreen will warn on\n\u003e first launch — click **\"More info\" → \"Run anyway\"** (signed builds land\n\u003e with beta).\n\u003e\n\u003e The desktop app **embeds the Praxia server inside the installer** —\n\u003e install, launch, paste an LLM provider key, and you're running. No\n\u003e separate `praxia serve` process to start, no `pip install`, no Python\n\u003e on the user's machine. Settings exposes only the three things a user\n\u003e actually controls: LLM provider keys (Anthropic / OpenAI / Google /\n\u003e Azure OpenAI / Qwen DashScope / Hugging Face — Gemma covered via all\n\u003e three cloud paths), local LLM (Ollama URL + model), and optional SSO\n\u003e tenant URL for org connection. Everything else (port, API key,\n\u003e storage layout, CORS) is managed by the app.\n\u003e\n\u003e **Multi-user organizational deployment** still works the same way:\n\u003e install Praxia on a shared host as `praxia serve`, point a custom\n\u003e frontend or SDK consumer at it, and several users share L3\n\u003e organizational memory + the L4 frozen layer with SSO / RBAC /\n\u003e audit / KMS-encrypted OAuth tokens — all in the OSS core.\n\u003e\n\u003e **Desktop-only features (in addition to everything the server offers):**\n\u003e\n\u003e - **🗂 Local folder ingestion with auto-discovery** — point Praxia at a\n\u003e   folder on your machine (e.g. `~/Documents/Contracts/`); the desktop app\n\u003e   walks it recursively, parses every supported file (PDF / DOCX / PPTX /\n\u003e   XLSX / TXT / MD / code), and makes the contents searchable by the agent\n\u003e   alongside L1 / L3 / L4 memory. Skips files above a size cap, watches\n\u003e   for new / changed files, re-indexes incrementally by mtime + content\n\u003e   hash. Useful for confidential documents you don't want uploaded to\n\u003e   cloud storage. (🚧 Phase 1b)\n\u003e - **🔔 Native notifications** when a long-running agent task finishes (🚧 Phase 1b)\n\u003e - **🪟 Native file dialogs** for drag-and-drop attachments\n\u003e\n\u003e *Cross-device continuity (PC ↔ phone handoff) and a mobile companion land\n\u003e in Phase 1b / Phase 2.*\n\u003e\n\u003e ---\n\u003e\n\u003e For library / SDK / CLI use, the Python install below is the developer path.\n\n```bash\n# 1. Install (pick the extras you actually need)\npip install praxia                              # Core\npip install \"praxia[ui,connectors,office,audio]\" # Common stack\npip install \"praxia[all]\"                       # Everything\n\n# 2. Configure once — all keys live in one place\npraxia config init      # interactive walkthrough\npraxia config show      # display resolved config (secrets masked)\npraxia config path      # show key resolution order\n# Or: cp .env.example .env  and edit\n\n# 3. Initialize\npraxia init --backend json --model auto\n# ↑ also bootstraps an `admin` user and writes its API key (one-time only)\n#   to `.praxia/auth/BOOTSTRAP_API_KEY.txt`. Save it somewhere safe and\n#   delete the file. Lost it? `rm -rf .praxia/auth \u0026\u0026 praxia init` issues\n#   a fresh key. See docs/quickstart.md § 3a for the full first-login flow.\n\n# 4. Run a flow (auto-parses .pdf / .docx / .xlsx / .pptx if attached)\npraxia run sales --customer-name \"Acme\" --product \"BizFlow\"\npraxia run logic --document spec.pdf\npraxia run rag --question \"What license is Praxia released under?\"\n\n# Run a business skill\npraxia skill run investment \"Mid-term thesis on a hypothetical mid-cap electronics issuer\"\npraxia skill run legal \"Review the risk in this services agreement\"\n\n# Launch the UI — fixed top-bar nav: Run / Prompts / Data / Knowledge /\n# Dashboard / Preferences (+ Admin for admin role). Login: User ID +\n# Password (= API key from `praxia user create`), or \"Sign in with\n# \u003cprovider\u003e\" via OIDC SSO when PRAXIA_SSO_PROVIDER is configured.\n# Cookie-based session survives reloads.\npraxia ui --port 8501\n\n# Personal → org memory distillation\npraxia consolidate --dry-run\npraxia freeze --block team_norms\n\n# Dashboards\npraxia dashboard --scope personal --user-id alice\npraxia dashboard --scope org\n\n# Admin: user management\npraxia user create alice --role member\npraxia user update alice --role operator --email alice@a.test\npraxia user deactivate alice\npraxia user delete alice --yes\npraxia user audit --limit 100\n\n# Admin: resource access policies (ACL — for IS depts)\npraxia policy add deny connector \"box:/Confidential/*\" \\\n    --principals \"role:member,role:viewer\" \\\n    --description \"Lock Confidential folder to operators+\"\npraxia policy list\npraxia policy test alice member connector box:/Confidential/q3.pdf read\n\n# Admin: data exports (CSV / JSON / JSONL — every export audit-logged)\npraxia admin export-audit audit.csv --since-days 30\npraxia admin export-users users.json --format json\npraxia admin export-memory ./memory_backup --all\npraxia admin export-policies policies.json\n\n# External connectors (Pull / Push, subject to ACL)\npraxia connector list\npraxia connector pull box 0 --limit 20 --save-to ./box_pulled\npraxia connector push salesforce Lead lead.json\npraxia connector pull kintone \"42?status='open'\"\n\n# Custom prompts (per-user + admin distribution)\npraxia prompt create my_qualifier prompt_body.txt\npraxia prompt list\npraxia prompt distribute curated_prompt body.md --target-roles member\n\n# Skill registry — promotion and admin distribution\npraxia skill promote --candidates\npraxia skill distribute investment_analyst --target-roles member,operator\n```\n\nMinimal Python example:\n\n```python\nfrom praxia import Praxia\nfrom praxia.flows import SalesAgentFlow\nfrom praxia.skills import InvestmentSkill\n\nm = Praxia(user_id=\"alice\", default_model=\"claude\")\n\n# Run a multi-agent flow\nresult = m.run(SalesAgentFlow, inputs={\n    \"customer_name\": \"Acme\",\n    \"product\": \"BizFlow\",\n})\n\n# Run a single business skill\nprint(InvestmentSkill().run(\"3-year investment thesis on Acme Mfg (TYO:0000)\"))\n\n# Personal memory accumulates automatically — no explicit save needed.\n# The nightly consolidator promotes effective patterns to org memory.\nm.consolidate(dry_run=True)\n```\n\nFull guide: [docs/quickstart.md](docs/quickstart.md).\n\n\u003e **Deploying it?** Two paths — fastest is `praxia ui` (full-stack); for \"Praxia as a brain behind your own frontend\" use the SDK or `praxia serve` (HTTP API). Setup recipes: [docs/deployment-modes.md](docs/deployment-modes.md).\n\u003e **Building a connector?** Step-by-step recipe in [docs/CUSTOM_CONNECTORS.md](docs/CUSTOM_CONNECTORS.md). The pattern is ~50 lines + an entry-point.\n\u003e **Formal specs?** Basic design / I/F / detailed design / **functional spec** (EN + JA) under [docs/specs/](docs/specs/).\n\u003e **Regression suite?** 364 tests covering auth/memory/exporters/CLI/i18n/etc. — see [docs/EVALUATION.md](docs/EVALUATION.md).\n\u003e **Multilingual?** Landing page + Streamlit UI ship in 8 languages (en / ja / zh-CN / ko / es / fr / de / pt-BR) with browser-language auto-detection — see [docs/i18n.md](docs/i18n.md).\n\u003e **Contributing?** PRs require a DCO sign-off (`git commit -s …`). Trademark policy + GDPR notes for operators are in [docs/legal/](docs/legal/).\n\u003e **MCP for Claude Desktop / Cursor?** Local stdio (`praxia mcp serve`) or remote HTTP+SSE (`/api/v1/mcp` after `praxia serve`). Every skill + flow becomes an MCP tool automatically.\n\u003e **OAuth scopes for connectors?** Per-provider scopes, app registration steps, least-privilege alternatives in [docs/OAUTH_SCOPES.md](docs/OAUTH_SCOPES.md).\n\u003e **Mobile-friendly?** Both the landing page and the Streamlit UI are responsive — chip-style nav on phones, scrollable tabs, ≥44px touch targets, compact mode toggle.\n\n---\n\n## 📐 Design Philosophy\n\n### 1. Capture tacit knowledge with **zero effort**\nNo explicit `CLAUDE.md`-style writing. Mem0/LangMem/HindSight extract entities and preferences from ordinary conversations.\n\n### 2. Promote only what's **effective**, **automatically**\nThree independent verdicts run in parallel. The framework auto-promotes only when consensus is high; medium-confidence items go to a review queue.\n\n### 3. Separate \"frozen\" from \"living\" knowledge\n- Living layer (shared blocks): updated instantly, all agents see it\n- Frozen layer (Markdown + git): only PR-reviewed, stable best practices\n\nThis keeps both **freshness** and **trust** intact.\n\n### 4. Use Graph storage **only where relationships are the value**\nMem0 OSS removed `graph_store` support in April 2026. We follow that signal: vector + entity linking is the default; graphs apply only to decision histories, customer 360, and incident causal chains.\n\n### 5. **Vendor lock-in is a non-goal**\n- LiteLLM lets any provider work\n- LTM backends are pluggable — and you can run several at once via [CompositeBackend / RoutedBackend](docs/FEATURES.md#51-multi-ltm-fusion--dynamic-routing-accuracy-boost) for higher recall without picking a winner\n- Markdown + git is the persistence layer of last resort\n- Apache 2.0 license, evolving toward an open-core model\n\n### 6. Ship \"**evidence**\" alongside the framework\nHallucination detection (`praxia.eval.hallucination`) and retrieval metrics (`praxia.eval.metrics`) are first-class. Customers don't have to take \"it works\" on faith.\n\nFor more, see [docs/design-philosophy.md](docs/design-philosophy.md).\n\n---\n\n## 📊 Use Cases by Industry\n\nDetailed Before/After tables for each domain are in **[docs/use-cases.md](docs/use-cases.md)**. Highlights:\n\n| Industry | Representative use case | Headline impact |\n|---|---|---|\n| Investment | Seed-stage VC due diligence | 4–6h → **45–60 min** per deck |\n| Sales | Pre-meeting research + storyboard | Proposal-acceptance rate **+15–20pt** |\n| Engineering Design | Requirements doc review | Senior architect time freed: **week 16h → 4h** |\n| Procurement | RFQ TCO comparison | Hidden costs found: **+30%** vs initial quote |\n| Patent | Prior-art search + novelty assessment | External patent-attorney fees **−50–70%** |\n| Legal | M\u0026A contract review | External law-firm costs **halved** (~$100k/deal) |\n\n**3-year compounding effects**: New-hire ramp **6–12mo → 2–3mo** / Veteran-departure knowledge loss **→ zero** / Cross-team best-practice diffusion **30+ items/month**.\n\n---\n\n## 🆚 When to pick what\n\nPraxia is opinionated for organizations that want **OSS + workflow templates\n+ auto personal-to-org memory cycling + integrated auth/ACL/audit** all in\none library. Adjacent tools have different goals — pick the one that fits\nyour need:\n\n- **[LangGraph](https://github.com/langchain-ai/langgraph)** — generic agent graph builder, fine-grained state machines, deep LangChain integration\n- **[CrewAI](https://github.com/crewAIInc/crewAI)** — lightweight role-based crew abstraction\n- **[AutoGen](https://github.com/microsoft/autogen)** — research-grade conversational multi-agents from Microsoft Research\n- **[Glean](https://www.glean.com/)** — hosted enterprise knowledge platform (no operational burden, commercial)\n- **[Mem0](https://github.com/mem0ai/mem0) / [LangMem](https://github.com/langchain-ai/langmem) / [Letta](https://github.com/letta-ai/letta) / [Zep](https://github.com/getzep/zep) / [HindSight](https://github.com/vectorize-io/hindsight)** — memory backends (Praxia uses them as plug-in backends, you can run several at once)\n\nThese are not mutually exclusive — Praxia uses Mem0 as a backend and can be\nembedded inside a LangGraph node.\n\nFor a feature-level matrix, see [`docs/COMPARISON.md`](docs/COMPARISON.md)\n(verifiable against each project's public documentation; corrections welcome\nvia Issues).\n\n---\n\n## 🗺 Roadmap\n\n| Phase | Scope | Status |\n|---|---|---|\n| **Phase 1** | Personal memory + 3 specialized flows + 6 business skills | ✅ **Done** |\n| **Phase 2** | Sleep-time consolidator + statistical (outcome-correlated) promotion | ✅ **Done** |\n| **Phase 3** | Shared blocks + Markdown freeze workflow + CLI | ✅ **Done** |\n| **Phase 4** | Skill registry promotion (personal → org) | ✅ **Done** |\n| **Phase 5** | Auth + RBAC + SSO + audit log + admin user CRUD | ✅ **Done** |\n| **Phase 5+** | Resource access policies (ACL) + admin data exports + custom prompts + 6 connectors + dashboards | ✅ **Done** |\n| **Phase 6** | Multi-tenant SaaS, advanced GUI, vertical editions | 🚧 Commercial |\n\n---\n\n## 🤝 Contributing\n\nWe're building a **community-driven library of industry recipes**. Three primary contribution paths:\n\n1. New workflow flows (`praxia/flows/`)\n2. New business skills (`praxia/skills/business/`)\n3. Industry recipes (`docs/recipes/`)\n\nSee [CONTRIBUTING.md](CONTRIBUTING.md).\n\n---\n\n## 📜 License\n\n[Apache License 2.0](LICENSE) — commercial use, modification, and redistribution permitted.\n\n**Copyright holder**: GenArch and Praxia Contributors.\n\nThird-party dependencies retain their own licenses; see [NOTICE.md](NOTICE.md) for the full attribution list.\n\n**Trademarks**: All product and company names referenced in Praxia documentation (Claude, ChatGPT, Gemini, Qwen, Box, SharePoint, Dropbox, Google Drive, kintone, Salesforce, Mem0, Letta, LangChain, CrewAI, Glean, etc.) are trademarks or registered trademarks of their respective owners. Praxia is not affiliated with, sponsored by, or endorsed by any of these companies — references are descriptive (nominative fair use) only. See [NOTICE.md § Trademark notice](NOTICE.md#trademark-notice) for the full list.\n\n**Demo data**: Company names in code examples (e.g., \"Acme Manufacturing\", \"AcmeAuto Inc.\") are **fictional** and for illustration only. Built-in skills include guardrails reminding users that final professional advice (investment, legal, patent, etc.) requires a qualified professional.\n\nWe may evolve toward an **open-core** model: enterprise GUI / advanced audit features under a separate license, while the framework remains Apache 2.0.\n\n---\n\n## 🚢 Deployment modes\n\nPraxia ships in two halves you can mix:\n\n| Mode | What you run | When to choose it |\n|---|---|---|\n| **A. Full-stack** | `praxia ui` (Streamlit) + Praxia core, one process | Internal team, fastest path |\n| **B-1. Embedded SDK** | Your Python service `import praxia` | You already have a Python backend |\n| **B-2. HTTP service** | `praxia serve` (FastAPI) + your own frontend | Non-Python frontend, mobile, or CDN-cached UI |\n\nBoth modes share the same auth, memory, and skills — only the frontend differs. Step-by-step setup, production checklist, and migration path: [docs/deployment-modes.md](docs/deployment-modes.md) ([JA](docs/deployment-modes.ja.md)).\n\n```bash\n# Full-stack\npraxia ui --port 8501\n\n# Backend-only HTTP API (8 endpoints under /api/v1)\npip install \"praxia[server]\"\npraxia serve --host 0.0.0.0 --port 8000 --cors-origin https://your-frontend.example\n```\n\n---\n\n## 📐 Design specs (formal documents)\n\nFor procurement / architecture review / extension work, formal design specs are available in **EN + JA**:\n\n| Document | English | JA |\n|---|---|---|\n| Basic design | [basic-design.en.md](docs/specs/basic-design.en.md) | [basic-design.ja.md](docs/specs/basic-design.ja.md) |\n| Interface spec | [interface-spec.en.md](docs/specs/interface-spec.en.md) | [interface-spec.ja.md](docs/specs/interface-spec.ja.md) |\n| Detailed design | [detailed-design.en.md](docs/specs/detailed-design.en.md) | [detailed-design.ja.md](docs/specs/detailed-design.ja.md) |\n| Functional spec | — | [functional-spec.ja.md](docs/specs/functional-spec.ja.md) |\n\n---\n\n## 🛠 Extending Praxia\n\nPraxia uses a **single extensibility primitive** (`praxia.extensions.Registry`) for every plugin point — connectors, memory backends, skills, flows, file parsers, output exporters, OAuth providers. Adding a plugin **does not require editing any core file**.\n\n| Plugin type | Base | Registry | Entry-point group | Lines |\n|---|---|---|---|---|\n| Connector | `Connector` protocol | `CONNECTORS` | `praxia.connectors` | ~50 |\n| Memory backend | `MemoryBackend` protocol | `BACKENDS` | `praxia.memory_backends` | ~80 |\n| File parser | `Parser` protocol | `PARSERS` | `praxia.parsers` | ~30 |\n| Output exporter | `Exporter` protocol | `EXPORTERS` | `praxia.exporters` | ~40 |\n| OAuth provider | `OAuthProviderConfig` | (instance) | `praxia.oauth_providers` | ~10 |\n| KMS adapter | `KmsAdapter` protocol | `KMS_ADAPTERS` | `praxia.kms_adapters` | ~30 |\n| Business skill | `Skill` | `SKILLS` | `praxia.skills` | ~20 |\n| Multi-agent flow | `Flow` | `FLOWS` | `praxia.flows` | ~30 |\n| Industry recipe | Markdown | n/a | — | n/a |\n\n**Custom connector tutorial** (end-to-end Notion example): [docs/CUSTOM_CONNECTORS.md](docs/CUSTOM_CONNECTORS.md) ([JA](docs/CUSTOM_CONNECTORS.ja.md)).\n\n**Two ways to register**:\n\n```python\n# (a) Decorator (in-tree contributions)\nfrom praxia.connectors.registry import CONNECTORS\n\n@CONNECTORS.register_decorator(\"notion\")\nclass NotionConnector: ...\n```\n\n```toml\n# (b) Entry-point (third-party packages — no fork needed)\n[project.entry-points.\"praxia.connectors\"]\nnotion = \"praxia_connector_notion:NotionConnector\"\n```\n\nAfter `pip install praxia-connector-notion`, the new connector shows up automatically in `praxia connector list`, the Streamlit UI, and the SDK — with **no edit to Praxia itself**.\n\nFull guide with examples for all 4 plugin types: **[docs/PLUGINS.md](docs/PLUGINS.md)**.\n\n---\n\n## 📈 ROI estimate (100-knowledge-worker mid-cap)\n\n| Variable | Year 1 | Year 2 |\n|---|---|---|\n| Workers in scope (N) | 100 | 100 |\n| Loaded cost / FTE (C) | $90k | $90k |\n| Routine work share (t) | 40% | 40% |\n| Time savings (s) | 35% | 60% |\n| Quality lift (Q) | $65k | $200k |\n| Praxia cost (P) | $80k | $80k |\n| **Net benefit** | **$1.25M** | **$2.30M** |\n\n3-year cumulative net ≈ **$5.2M**. Even halving every parameter still produces \u003e 10× ROI.\n\nFull model + worked examples: [docs/FEATURES.md#roi-projection-model](docs/FEATURES.md#14-roi-projection-model).\n\n---\n\n## 📚 Acknowledgements \u0026 Inspirations\n\n- [Mem0](https://github.com/mem0ai/mem0) — personal memory layer\n- [Letta](https://github.com/letta-ai/letta) — shared memory blocks concept\n- [LangMem](https://github.com/langchain-ai/langmem) — long-term memory SDK\n- [Zep](https://github.com/getzep/zep) — temporal knowledge graph for agent memory\n- [HindSight](https://github.com/vectorize-io/hindsight) — Experience / Entity Summary / Belief model\n- [LiteLLM](https://github.com/BerriAI/litellm) — unified provider abstraction\n- [Claude Skills](https://docs.claude.com/) — skills registry conventions\n- [Model Context Protocol](https://modelcontextprotocol.io) — tool/skill interop\n\nTheoretical groundwork:\n- [LinkedIn Cognitive Memory Agent](https://www.infoq.com/news/2026/04/linkedin-cognitive-memory-agent/) (Episodic + Semantic + Procedural)\n- [Mem0 paper](https://arxiv.org/abs/2504.19413) (arXiv:2504.19413)\n- [Letta sleep-time agents](https://docs.letta.com/guides/agents/architectures/sleeptime/)\n\n---\n\n\u003e **Mission**: Bridge \"individual brilliance\" and \"organizational continuity\" with AI.\n","project_url":"https://awesome.ecosyste.ms/api/v1/projects/github.com%2Fpraxia-dev%2Fpraxia","html_url":"https://awesome.ecosyste.ms/projects/github.com%2Fpraxia-dev%2Fpraxia","lists_url":"https://awesome.ecosyste.ms/api/v1/projects/github.com%2Fpraxia-dev%2Fpraxia/lists"}