Turn any AI agent into a
verifiable domain expert

Structured knowledge packs on Schema v4.1 — one concept = one retrieval atom — served through ExpertPack MCP. Provenance-first grounding, reconstructable citations, and measurable retrieval quality.

Get EP MCP Browse schema & packsSee How It Works

Give your AI the knowledge it's missing

Esoteric knowledge (EK) is knowledge not found in the weights of frontier LLMs. It's the tribal knowledge in your support team's heads, the gotchas your engineers learned the hard way, the decision patterns your founder never wrote down — the gap between what a model can answer and what an expert actually knows.

ExpertPacks deliver this knowledge to any AI agent in a way that minimizes token cost and maximizes prompt quality through RAG. Every pack is built from atomic-conceptual concept files, exported through a compact Agent Knowledge Schema, and measured by its EK ratio — the proportion of content that frontier models cannot correctly produce on their own. During hydration, every fact is triaged: esoteric knowledge gets maximum treatment, general knowledge gets compressed to scaffolding. The result is dense, high-value context that makes your AI genuinely expert — not just articulate.

Built on ExpertPack Schema 4.1 — one concept = one retrieval atom, requires: expansion, content_hash provenance.

🧠

EK-Optimized

Every fact is triaged during hydration — maximize esoteric knowledge, compress what models already know

📊

Measurable Quality

EK ratio, correctness, hallucination rate, and refusal accuracy — measured, not guessed

🎯

Atomic-Conceptual Design

Self-contained concept files. One concept per file, each carrying its own definition, body, FAQs, and related terms — authored as a single retrieval unit so what you write is what the agent sees.

📝

Markdown-First

Human-readable, AI-consumable, git-versionable — no proprietary formats or lock-in

⚡

Token-Efficient

Three-tier context strategy loads only what's needed per turn

🧾

Verifiable Grounding

Stable IDs, content hashes, verification dates, and reconstructable source spans make answers auditable — not just plausible

"Why can't my AI just search for this?"

Three reasons web search can't replace an ExpertPack.

🤥

Models don't know what they don't know

When a model confidently hallucinates, it doesn't trigger a search. It doesn't think "I'm unsure, let me look this up" — it thinks it already knows. An ExpertPack loaded into context preempts the hallucination with the correct answer before the model gets a chance to fabricate.

🔍

Search requires the right question

Even with tool-use, the model needs to know what to search for. If it doesn't know about a specific firmware bug, it won't search for the precise query that finds the fix — it'll search generically and get generic results. You can't search for knowledge you don't know exists.

🔒

Not all knowledge is on the internet

Source code analysis reveals undocumented behavior that exists nowhere online. Expert interviews capture tribal knowledge that was never written down. Person packs contain private stories and reasoning. These are original knowledge sources — no search engine indexes them.

Three pack types, infinite use cases

🧑

Person Packs

Capture a person — stories, beliefs, relationships, voice, and legacy.

Use cases: Personal AI assistant, family archive, memorial AI, digital legacy, founder knowledge capture
📦

Product Packs

Deep knowledge about a product or platform — concepts, workflows, troubleshooting.

Use cases: AI support agent, sales assistant, training tool, onboarding guide, product documentation
🔄

Process Packs

Complex multi-phase processes — phases, decisions, checklists, gotchas.

Use cases: Home building guide, business formation, project management, certification processes
🔗

Composites

Combine multiple packs into a single agent deployment with role assignments and context control.

Use cases: CEO AI assistant, multi-product support bot, company knowledge base, personal legacy AI

How it works

1

Point your AI at the schema

Pick a pack type — person, product, or process. Your AI agent reads the schema and knows exactly what to build.

2

Feed it knowledge

Talk to the agent, point it at websites, drop in documents, or hand it data exports. It structures everything automatically.

3

Deploy the pack

Drop the pack into any AI agent's workspace, or serve it through EP MCP (ep_search / ep_read). Instant domain expertise — no prompt engineering required.

4

Measure & improve

Run evals to measure correctness, completeness, and hallucination rate. Use results to guide targeted improvements.

Free community packs

Open-source ExpertPacks built from real documentation, community forums, and source code analysis. Each pack shows its EK ratio — the percentage of content that frontier AI models cannot produce on their own. Higher EK = more value your AI can't get anywhere else. Download individual packs directly from GitHub — no account required. ⭐ Star the repo if you find them useful!

🏠

Home Assistant

Composite PackEK 54%

The open-source home automation platform. Deep practitioner knowledge covering smart home protocols, automation patterns, presence detection, YAML configuration, ESPHome, dashboards, voice assistant, energy management, and security monitoring. Includes community-sourced gotchas and real-world device compatibility data.

📄 132 files📏 600 KB📝 12,200+ lines
Zigbee / Z-Wave / MatterAutomationsPresence DetectionESPHomeDashboardsVoice AssistantEnergy
DownloadBrowse
🎨

Blender 3D

Product PackEK 42%

The free, open-source 3D modeling, animation, and rendering software used by millions of artists and studios worldwide. Covers polygon modeling, sculpting, animation & rigging, physics simulation, PBR shading, Cycles/EEVEE rendering, Geometry Nodes, compositing, Python scripting, and production workflows.

📄 83 files📏 392 KB📝 7,200+ lines
Modeling & TopologyAnimation & RiggingSculptingShading & PBRCycles / EEVEEGeometry NodesPhysics & SimulationCompositingPython ScriptingGame ExportProduction Workflows
DownloadBrowse
☀️

Solar & Battery DIY

Composite PackEK 52%

A practitioner guide for residential solar panel and battery storage systems. Covers system design, panel and battery product comparisons, NEC code compliance, permitting, installation best practices, and troubleshooting.

📄 53 files📏 291 KB📝 3,600+ lines
System DesignPanel SelectionBattery StorageNEC CodePermittingTroubleshooting
DownloadBrowse

Grounded retrieval you can verify (Schema v4.1).

Basic RAG embeds documents and retrieves top-k chunks. ExpertPacks author each concept as a single self-contained file, declare dependencies explicitly, and preserve provenance all the way to runtime. With Reconstruct Mode, an agent can show the exact Markdown span, byte offset, and content_hash / SHA-256 proof that grounded the claim.

📦

Atomic-Conceptual Concept Files

Self-contained retrieval units. One file per concept carries its definition (opening paragraph), body explanation, FAQs, and related terms — authored as what the agent actually sees. concepts/territory.md (live example).

🔗

Declared Dependencies (requires:)

When an overview atom needs a detail atom to be fully useful, authors declare it in frontmatter: requires: [detail-atom]. The retrieval layer auto-appends the required atoms when the overview is hit — directional, transitive, and bounded by depth and token caps so it never displaces primary results.

📐

Right-Sized Atoms

Soft target 500–800 tokens per concept, hard ceiling at 1,000. Oversized concepts split into independent atoms connected by requires: rather than hierarchical file groups. Every split is still a first-class retrieval unit on its own.

🧬

Agent Knowledge Schema

Export compact AKS JSONL rows — id, canonical_statement, canonical_path, source_checksum, and graph edges — for token-efficient retrieval pipelines.

🔍

Reconstruct Mode

Reconstruct Mode returns the original Markdown span, byte offset, and content_hash / SHA-256 proof with every hit. Trace any answer back to exactly what grounded it — stable ID, source file, and verification date included.

🕸️

Ontology-Aware Graphs

Suggest entities from pack content, accept them into ontology.yaml, and export ontology/entity edges into _graph.yaml for richer GraphRAG traversal.

Works where your agents already run

Open-source tooling for building, measuring, and deploying ExpertPacks — across the platforms you already use.

OpenClaw

OpenClaw

Battle-tested with OpenClaw. Add the pack path to memorySearch.extraPaths — instant expertise.

Claude

Claude Cowork

Load packs as workspace Markdown, or connect via MCP for live ep_search / ep_read.

Microsoft Copilot

Microsoft Copilot

Connect through MCP or custom retrieval so Copilot agents query ExpertPack atoms on demand.

Grok Bot

Grok Bot

Use MCP tools or drop packs into workspace Markdown — same Schema v4.1 atoms either way.

📐

Retrieval-Ready by Design

Files are authored as self-contained retrieval units (soft 500–800 tokens, hard ceiling 1,000). Any RAG chunker passes them through intact — no external tooling needed. The schema IS the chunking strategy. Chunk sidecars preserve provenance across pipelines.

📊

EK Ratio Measurement

Blind-probes frontier models to measure what % of your pack they can't produce alone.

🧪

Eval Runner

Automated eval execution with LLM-as-judge scoring for correctness, hallucination, and refusal.

🧬

AKS Export

Emit compact Agent Knowledge Schema JSONL and run strict export-readiness gates for CI and retrieval pipelines.

🕸️

Ontology Suggest

Generate review-first entity/category suggestions, accept them into ontology.yaml, then export ontology-aware graph edges.

✅

Validator & Doctor

Run ep-validate --strict and ep-doctor to catch broken requires: links, token overruns, and frontmatter issues before retrieval does.

Connect with EP MCP →

Stop burning tokens on context bloat

Generic RAG dumps documents into a vector store and loads everything into context — hoping the model will sort it out. You pay for every irrelevant token on every turn.

ExpertPacks use a three-tier context strategy: core identity loads every session, knowledge loads on topic match, and heavy content loads only on demand. Your agent gets the right information at the right time — not everything all the time.

  • Token costTiered loading — only pay for what this turn actually needs
  • RetrievalAtomic concept files + declared requires: dependencies + graph traversal
  • StructureSchemas model real expertise — not just document chunks
  • QualityEval framework measures correctness and catches hallucinations
  • ProvenanceStable IDs, verification dates, checksums, and reconstructable source spans
  • DependenciesDeclared requires: links between atoms — honored at retrieval time
  • CompositionCombine packs — person + product + process in one agent
  • PortabilityPlain Markdown — works anywhere, version-controlled

Serve packs as an API — ExpertPack MCP

ExpertPack MCP (EP MCP) is an open-source MCP server that turns any ExpertPack into a live, queryable knowledge service. Hybrid BM25 + vector retrieval, requires: expansion, Reconstruct Mode, and multi-pack serving — with a live EDEP eval at 97.6% source hit rate. Connect OpenClaw, Claude Cowork, Microsoft Copilot, Grok Bot, Claude Desktop, Cursor, Windsurf, or any MCP client (HTTP + stdio).

🔍

Hybrid Retrieval

BM25 + vector search (sqlite-vec), metadata boosting, MMR, requires: expansion, and ontology-aware graph traversal. Live EDEP eval: 97.6% source hit rate, 0 full misses.

📁

EP-Native Chunking

Files are atomic retrieval units — no arbitrary splitting. Frontmatter drives filtering, AKS export, graph edges, chunk sidecars, and provenance metadata that flows through to every result.

🔍

Reconstruct Mode

Opt in to original Markdown spans, byte offsets, and content_hash / SHA-256 verification so agents can prove which source text grounded an answer.

📚

Multi-Pack

Serve many packs from one host via /packs/{slug}/mcp (HTTP) or stdio pack selection. Switch domains without redeploying the agent.

🔌

Any MCP Host

Streamable HTTP (cloud-ready) + stdio (local). Works with OpenClaw, Claude Cowork, Microsoft Copilot, Grok Bot, Claude Desktop, Cursor, Windsurf, and any MCP client.

⚡

Self-Hostable

Point it at any ExpertPack with a single config file. Runs locally or in the cloud — no vendor lock-in, no external dependencies beyond your embedding provider.

Get EP MCP on GitHub →

Built for serious knowledge engineering

📊

Evaluation Framework

Standardized eval sets measure correctness, completeness, hallucination rate, and refusal accuracy — including Typed Answer Contract (TAC) checks. Run automated evals with the included eval runner. Track quality over time with baselines and scorecards.

🏷️

Validator & Doctor

ep-validate --strict and ep-doctor check pack structure, frontmatter, token budgets, chunk sidecars, and cross-atom requires: references. Catch broken dependencies and oversized atoms before retrieval does.

📖

Guides & Tooling

Population guides cover every knowledge source. Validator, AKS export, ontology suggest, graph export, and eval runner turn pack building into an auditable engineering workflow.

💎

Obsidian Compatible

Every ExpertPack is a valid Obsidian vault. Open any pack in Obsidian and get live Dataview queries by content type, EK score, and tags — graph view, full-text search, and template-based authoring included. Standard Markdown links keep packs fully readable on GitHub and in any editor simultaneously.

NEW

Export your AI agent as an ExpertPack

Your agent accumulates months of knowledge — identity, preferences, infrastructure expertise, behavioral patterns, relationships. Distill that into a portable ExpertPack and bootstrap a new instance on any platform that can read Markdown or speak MCP.

🔍

Auto-Discover

The agent scans its own workspace, classifies every knowledge chunk, and proposes constituent packs — agent, person, product, process.

⚗️

Distill

Raw state (journals, configs, memory files) is compressed into structured, deduplicated EP-compliant files. 438KB raw → 31KB distilled.

📦

Package

A composite EP wires the agent pack (voice) with person/product/process packs (knowledge). Ready for any platform that can read Markdown or speak MCP.

💾

Backup & Restore

Your agent dies — spin up a new one from its EP. Immediately competent, not starting from scratch.

🚚

Platform Migration

Move from one AI platform to another. Your agent's knowledge comes with it — portable by design.

🤝

Agent Collaboration

Share domain expertise between agents. One agent's product knowledge becomes another's via composite.

🏪

Marketplace Ready

Distribute well-trained agent configurations as portable packs. Built-in privacy controls keep secrets out.

Works with your stack
OpenClawClaude CoworkMicrosoft CopilotGrok Bot

Battle-tested with OpenClaw; also used with Claude Cowork, Microsoft Copilot, and Grok Bot — plus any MCP client.

Start building on Schema v4.1

Open source. Apache 2.0. Free forever. Connect agents with EP MCP.

Get EP MCP View framework ⭐ Star

If ExpertPack is useful to you, a GitHub star helps others discover it.