Build Context Packs
Turn local files, Markdown records, source maps, validation rules, and export profiles into reusable AI-ready context.
Context layer / source-backed / ready for agents
Validated context. Ready for agents.
Contextarr is the context layer: a self-hosted system that turns local files, records, sources, and rules into validated Context Packs for AI assistants and agents. It is not an agent runner.
- Source-backed
- Structured
- Validated
- Agent-ready
- Developer-first
- Privacy-first
Core object
Context Packs are the unit of reuse.
A Context Pack is a structured, versioned, source-backed bundle of reusable context for humans and AI tools.
Examples
- AI Workstation Pack
- Jellyfin Server Pack
- Claude Code Project Pack
- Internal Support KB Pack
- Fake Product Line Pack
{
"pack_id": "ai-workstation-pack",
"version": "0.7.0",
"records": ["records/*.md"],
"source_map": "sources.json",
"export_profiles": ["chatgpt", "claude", "codex", "json_records"],
"review_status": "human-reviewed"
}Review and validate
Catch missing fields, stale records, broken source references, sensitive exports, and unreviewed drafts before context reaches an AI tool.
Export anywhere
Generate profile-driven context for ChatGPT, Claude, Codex, Claude Code, generic Markdown, JSON records, and read-only MCP clients.
How it works
One local pipeline from source material to AI-ready context.
Your source files stay local. SQLite is a rebuildable index. Exports are generated artifacts. MCP is read-only.
- 01 Local files
Markdown, JSON, source maps, rules, and profiles stay on disk.
- 02 Validate
Schemas and deterministic checks catch missing or unsafe context.
- 03 Index
SQLite is rebuilt from packs and remains derived local state.
- 04 Review
Health and review queues keep draft context visible.
- 05 Render
Records become sanitized Markdown or static HTML.
- 06 Export
Profiles generate target-ready artifacts with redaction rules.
- 07 Read-only MCP
Clients can inspect approved context without mutating files.
Current build status
Early software, real implementation.
Contextarr is not production ready. The public repo now has the local Context Pack workflow implemented through Phase 11, while Phase 12 is documentation and terminology planning for future Skills and Agent Kits only.
Current status Phase 12: terminology planning
No Skill schema code, Agent Kit runtime, cloud sync, marketplace, or agent runner.
Implemented
- TypeScript monorepo
- Zod schemas
- read-only validator
- demo Context Packs
- SQLite derived index
- Fastify local API
- React/Vite dashboard
- sanitized Markdown renderer
- static HTML renderer
- Pack Health
- Review Queue
- export engine
- Export Center
- local CLI/core importers
- read-only Composer previews
- read-only stdio MCP server
- Docker Compose local preview
- Phase 12 terminology docs for future Skills and Agent Kits
Not included yet
- web importer UI
- API import endpoints
- pack file editing from review actions
- saving composed exports as new packs
- Skill schema and Skill validator
- demo Skills or demo Agent Kits
- Skill Library or Agent Kit UI/API/MCP extensions
- Skill execution or Agent Kit runtime behavior
Security boundaries
Built around hard boundaries.
Contextarr prepares local, reviewable context. It does not host a cloud vault, phone home, execute pack content, or run agents.
- No hosted cloud
- No telemetry
- No executable packs
- No scripts inside packs
- No shell commands
- No hidden network calls
- No direct Gmail, bank, or brokerage connectors
- No agent runner
- Read-only MCP
- AI-drafted content requires human review
Future direction
Skills and Agent Kits come after the pack system is stable.
This is planned after the core Context Pack system is stable. Contextarr will prepare Agent Kits, not run them.
Skills tell agents how to work.
Context Packs tell agents what to know.
Agent Kits combine both for a specific task.