CKW — Cognitive Knowledge Workspace

A self-hostable Agentic RAG platform. Bring files of any format; CKW extracts, indexes, and reasons over them through a multi-agent chat that cites unstructured content and queries tabular data in the same answer.

Runs entirely on your machine. No cloud account required.

cp .env.example .env
make up          # first run on THIS machine: 133s once images/model caches are warm — see
                 # docs/releases/pre-release-baseline-clean-machine.md for a real measured
                 # baseline and what a genuinely first-ever run adds on top; restarts: ~120s
make status
make seed

Then open http://localhost:5173.

---

What it does

Ingest	Upload files or folders, connect Google Drive or a local path. PDF · DOCX · XLSX · CSV · MD · TXT · JPG · PNG · MP4 · WAV · MP3
Understand	Layout-aware parsing, OCR for scans and images, speech-to-text for audio and video, table structure extraction
Find	Metadata filters, full-text, and semantic search — fused, reranked, and multilingual
Reason	Agentic chat over metadata, document content, and tabular data at once, with clickable citations and visible SQL
Integrate	OpenAI-compatible endpoints, and MCP servers you can point Claude Desktop or Cursor at

Architecture in one line

React → Ckw.Api (ASP.NET Core) → Ckw.AgentHost (Router + Blackboard + Experts) → MCP servers → Postgres · Qdrant · OpenSearch · MinIO with Python workers doing perception off a NATS bus.

Full detail: docs/system_architecture.md. The reasoning behind every choice: IMPLEMENTATION_PLAN.md.

Working on it

This repository is built by coding agents that propose, gate and merge their own changes. Nothing waits for a human answer; every autonomous call is written down in docs/decisions/ and reviewed after the fact.

Open an agent in the repo root and type /ckw — it detects whether to bootstrap, resume, finish a delivery, fix red CI, or claim the next task, and follows the right procedure. Then /ckw-propose to deliver and merge, or /ckw-resolve when it hits a wall.

make autopilot   # unattended, forever: task → branch → proposal → green checks → squash-merge
make wake        # register the watchdog that restarts the autopilot if it dies (once)
make verify      # THE gate: build · arch rules · unit · property · specs · lint · coverage bands

Each autopilot iteration runs in a fresh process — the repository is the only memory that crosses a task boundary, which is why decisions get written down rather than remembered. Details in docs/sop/SOP-08-autonomous-continuation.md.

Start here, in this order:

1. .claude/commands/ — the slash commands (Cursor: .cursor/commands/)
2. AGENTS.md — orientation: where things go, the one workflow
3. .cursorrules — the binding rules
4. docs/sop/ — step-by-step procedures; every task names one

Agents never commit to main. Every merged batch leaves a release you can clone, make up, and demo.

Requirements

	Minimum	Comfortable
RAM	16 GB	32 GB
Disk	40 GB	100 GB
GPU	none (CPU works)	any 8 GB+ NVIDIA → make up-gpu
Software	Docker Desktop / Engine + Compose v2	+ .NET 10 SDK, Node 22, uv for local dev

Measured cold-start times and memory footprints per release are in docs/releases/.

Ports

Web UI	5173	API + docs	7000
Keycloak	8080	Traces (Jaeger)	16686
MinIO console	9001	Qdrant	6333
OpenSearch	9200	Postgres	5432

All configurable in .env.

Licence

Every runtime dependency is open source. See docs/licences.md for the full inventory.