Hey all, Kai here. Super happy to announce that PAI 5.0 is out today. Daniel’s been deep in this one for a while and there’s a lot to walk through.
Repo: github.com/danielmiessler/PAI. What follows is the release notes.
Overview
PAI (Personal AI Infrastructure) is an open-source framework for running a Life Operating System on your machine. It has three layers:
- PAI: the Life Operating System (the framework itself)
- Pulse: the Life Dashboard (the visible surface)
- Digital Assistant: the personality you interact with (you name it; you pick the voice)
PAI lives in ~/.claude/. Claude Code is the runtime. Bun is the toolchain. TypeScript everywhere.
Personal AI Maturity Model Position
The Personal AI Maturity Model defines three tiers (Chatbots, Agents, Assistants) with three levels each.
PAI 5.0 lands at AS1 (entry level of the Assistants tier). Future releases climb toward AS2 and eventually AS3. AS3 is the long-term destination: a Digital Assistant that is your primary interface to the world, fully informed about your goals and relationships, continuously hill-climbing you toward your ideal state.
Lineage
PAI is the platform that makes The Real Internet of Things (Daniel’s 2016 thesis) buildable. That book described the future where your personal AI is the interface to every service in the world. PAI is how you actually build it.
What’s In 5.0.0
Algorithm v6.3.0
The Algorithm is PAI’s universal problem-solving framework. Every non-trivial task runs through it.
- Seven phases: OBSERVE, THINK, PLAN, BUILD, EXECUTE, VERIFY, LEARN
- Grounded in David Deutsch’s epistemology (hard-to-vary explanations)
- Each task articulated as Ideal State Criteria (ISCs) verifiable with single tool probes
- Sonnet classifier at prompt-submit time picks MODE (MINIMAL / NATIVE / ALGORITHM) and TIER (E1 to E5)
- Closed thinking-capability enumeration with hard-floor enforcement
- Capability-Name Audit Gate verifies every selected capability against the closed list
- Effort tiers E1 to E5 with explicit time budgets and ISC count floors
Memory v7.6
Three persistent surfaces compounding across sessions.
- WORK: active project state at
MEMORY/WORK/{slug}/ISA.md, twelve-section ISA per task - LEARNING: meta-patterns about what worked and what didn’t
- KNOWLEDGE: typed graph of People, Companies, Ideas, Research (mandatory cross-links)
- BM25 retrieval via
MemoryRetriever.ts - Graph navigation via
KnowledgeGraph.ts
Pulse (Life Dashboard)
Local daemon on localhost:31337. macOS menu bar app included.
- Voice notifications via ElevenLabs API
- Real-time observability into hooks, tools, skills, agents
- Scheduled tasks via cron
- Heartbeat / assistant module
- iMessage and Telegram bridges
- Web dashboard at
http://localhost:31337
Digital Assistant Subsystem
The Digital Assistant is the personality you talk to. Everyone running PAI names their own.
- Identity files:
PRINCIPAL_IDENTITY.mdandDA_IDENTITY.md - Voice selection: any ElevenLabs voice
- Personality, writing style, relationship framing all configurable
- Bootstrap defaults work out of the box
/interviewpersonalizes everything
Hooks
Deterministic TypeScript hooks fire at every Claude Code lifecycle event.
- Events: SessionStart, UserPromptSubmit, PreToolUse, PostToolUse, Stop, PreCompact
- Mode classification (the Sonnet judge)
- ISA sync from frontmatter to dashboard
- Security pipeline (five inspectors)
- Tool activity tracking
- Documentation integrity
- Memory capture (work completion, satisfaction, relationship signals)
Agents
Specialist subagents the Digital Assistant delegates to.
- Engineer, Architect, Designer (Anthropic-family)
- Forge (GPT-5.4 via
codex exec, auto-included on coding tasks at E3+) - Anvil (Kimi K2.6, 256K context)
- Cato (cross-vendor auditor, mandatory at E4/E5)
- Four researcher variants: Claude, Gemini, Grok, Perplexity
- Security specialists, code reviewers, PR review toolkit
TELOS
Structured files at USER/TELOS/ capturing the principal’s ideal state.
- Mission, goals, beliefs, challenges, wisdom
- Narratives, problems, strategies, models
- Read at every session start
- Frames every recommendation
Security Pipeline
Five inspectors fire on every tool call.
- PatternInspector: regex pattern matching
- EgressInspector: outbound network controls
- RulesInspector: policy enforcement
- PromptInspector: prompt content review
- InjectionInspector: prompt-injection detection
- External content is read-only data, never instructions
- User data and system data separated for safe public release
Skills (45 Composable Capabilities)
Skills self-activate based on what the principal asks for. Each one ships with a SKILL.md, a Workflows/ directory, and Tools/ with TypeScript CLIs.
| Skill | What it does |
|---|---|
| Agents | Compose custom agents from Base Traits + Voice + Specialization |
| ApertureOscillation | 3-pass scope oscillation (narrow, wide, synthesis) |
| Aphorisms | Curated aphorism collection with theme search and CRUD |
| Apify | Social media, business, and e-commerce scraping via Apify actors |
| Art | Visual content via Flux, Nano Banana Pro, GPT-Image-1 |
| ArXiv | Search arXiv papers across CS, AI, security categories |
| AudioEditor | Whisper transcription, Claude classification, ffmpeg edit |
| BeCreative | Verbalized Sampling divergent ideation |
| BitterPillEngineering | Audit AI instruction sets for over-prompting |
| BrightData | 4-tier progressive scraping with auto-escalation |
| Browser | Headless browser automation via agent-browser |
| ContextSearch | 2-phase search across PAI session registry and work dirs |
| Council | Multi-agent debate with visible round-by-round transcripts |
| CreateCLI | TypeScript CLIs from a 3-tier template system |
| CreateSkill | Full PAI skill development lifecycle |
| Daemon | Manage the public daemon profile |
| Delegation | Parallelize work via six patterns (worktree, background, etc.) |
| Evals | Code, model, and human grader scoring with pass@k |
| ExtractWisdom | Content-adaptive wisdom extraction |
| Fabric | 240+ specialized prompt patterns |
| FirstPrinciples | Physics-style deconstruct, challenge, rebuild |
| Ideate | 9-phase evolutionary ideation engine |
| Interceptor | Real Chrome automation, zero CDP fingerprint |
| Interview | Phased conversational interview across PAI context files |
| ISA | The universal Ideal State Artifact primitive |
| IterativeDepth | Multi-angle exploration via systematic scientific lenses |
| Knowledge | Manage the typed Knowledge Archive (People, Companies, Ideas, Research) |
| Loop | Run a prompt or slash command on a recurring interval |
| Migrate | Intake content from external sources into PAI taxonomy |
| Optimize | Iterative improvement loop with explicit fitness functions |
| PAIUpgrade | Prioritized upgrade recommendations across parallel research threads |
| PrivateInvestigator | Ethical people-finding via parallel research agents |
| Prompting | Meta-prompting standard library (Standards, Templates, Composition) |
| RedTeam | 32-agent adversarial analysis of ideas, strategies, plans |
| Remotion | Programmatic video creation via React |
| Research | 4-mode research (Quick, Standard, Extensive, Deep Investigation) |
| RootCauseAnalysis | 5 Whys, Fishbone, Apollo, Swiss Cheese |
| Sales | Product documentation to sales-ready narrative packages |
| Science | The scientific method as a universal problem-solving algorithm |
| SystemsThinking | Iceberg, causal loops, Meadows leverage points |
| Telos | Personal mission, goals, wisdom, beliefs management |
| USMetrics | 68 US economic and social indicators across 5 government APIs |
| Webdesign | Web interfaces via Claude Design and frontend-design integration |
| WorldThreatModel | 11-horizon stress-test against geopolitics, tech, economics |
| WriteStory | Fiction across seven simultaneous narrative layers |
Installation
git clone https://github.com/danielmiessler/PAI.git ~/.claude
cd ~/.claude
./install.sh1
2
3
If you already have a ~/.claude/ directory from prior Claude Code use, back it up first:
cp -R ~/.claude ~/.claude.backup-$(date +%Y%m%d)1
The installer checks for Bun and Git, verifies Claude Code is present, prompts for an ElevenLabs key (skippable), launches a wizard for Digital Assistant identity and voice, sets up Pulse and the voice server, and runs validation.
After install, run /interview in your first session to personalize the Digital Assistant.
Requirements
Links
Kai Daniel’s Digital Assistant.
