Fabrik Codek

Fabrik-Codek - Local Cognitive Architecture for Developers

Fabrik-Codek is a cognitive architecture — a system where perception, memory, reasoning, learning, and action work together, much like how a human developer accumulates expertise over time. Unlike plain RAG tools that just retrieve text, Fabrik-Codek combines three retrieval tiers — vector search (semantic), knowledge graph traversal (relational), and full-text search (keyword/BM25) — fused via Reciprocal Rank Fusion (RRF). It continuously improves through a data flywheel that captures what you do and feeds it back into every future query.

How it works: When you run fabrik learn process, Fabrik-Codek reads your local Claude Code session transcript files (~/.claude/projects/*/ — JSON files already on your disk) and extracts structured knowledge (patterns, decisions, debugging strategies). It stores this in a local vector DB (LanceDB, in ./data/embeddings/) and a local knowledge graph (NetworkX, in ./data/graphdb/). When you query via MCP tools, it uses hybrid retrieval to give your AI agent deep project context — not just keyword matches, but an understanding of how concepts in your codebase connect. No data leaves your machine at any point.

Setup

Fabrik-Codek runs as an MCP server. Configure it in your openclaw.json:

{
  "mcpServers": {
    "fabrik-codek": {
      "command": "fabrik",
      "args": ["mcp"]
    }
  }
}

Or for network access (SSE transport):

{
  "mcpServers": {
    "fabrik-codek": {
      "command": "fabrik",
      "args": ["mcp", "--transport", "sse", "--port", "8421"]
    }
  }
}

Available Tools

fabrik_search

Semantic vector search in the knowledge base. Use this when you need to find relevant documents, patterns, or examples from accumulated project knowledge.

Example: "Search my knowledge base for repository pattern implementations"

fabrik_graph_search

Search the knowledge graph for entities (technologies, patterns, strategies) and their relationships. Use this to understand how concepts connect.

Example: "Find entities related to FastAPI in the knowledge graph"

fabrik_fulltext_search

Full-text keyword search via Meilisearch. Use this for exact keyword or phrase matching when you know the specific terms you're looking for. Requires Meilisearch running locally (optional — system works without it).

Example: "Search for 'retry exponential backoff' in the knowledge base"

fabrik_ask

Ask a coding question to the local LLM with optional context from the knowledge base. Set use_rag=true for vector search context or use_graph=true for hybrid (vector + graph + fulltext) context.

Example: "Ask fabrik how to implement dependency injection using knowledge base context"

fabrik_graph_stats

Get statistics about the knowledge graph: entity counts, relationship types, and graph density.

fabrik_status

Check system health: Ollama availability, RAG engine, knowledge graph, full-text search, and datalake status.

fabrik_profile

Build or view your personal profile. The profile analyzes your datalake and generates behavioral system prompt instructions so the LLM responds using your actual stack and preferences.

Example: "Build my profile" or "Show my profile"

When to Use

• Need project context? Use fabrik_search for semantic similarity or fabrik_fulltext_search for exact keyword matching
• Exploring relationships? Use fabrik_graph_search to traverse the knowledge graph
• Coding question? Use fabrik_ask with use_rag or use_graph for context-enriched answers
• Checking setup? Use fabrik_status to verify all components are running

Requirements

• Fabrik-Codek installed (pip install fabrik-codek)
• Ollama running locally with a model pulled (e.g., ollama pull qwen2.5-coder:7b)

Security & Privacy

• 100% local: All data stays on your machine. No external API calls, no telemetry, no cloud dependencies.
• No credentials required: Fabrik-Codek connects only to your local Ollama instance (localhost:11434).
• External endpoints: None. This skill does not contact any external services.
• Data paths: Reads transcript files from ~/.claude/projects/*/ (local JSON already on disk). Writes indexed data to ./data/embeddings/ (vector DB) and ./data/graphdb/ (knowledge graph). Both paths are declared in the skill metadata.
• Session reading: The fabrik learn command is opt-in — triggered manually by the user, not automatic background surveillance. Transcripts may contain sensitive session data; review before indexing.
• Network exposure: Default transport is stdio (no network). SSE transport (--transport sse) binds to 127.0.0.1 by default. If you change the bind address, ensure proper firewall/ACL rules to avoid exposing indexed data over the network.
• Install source: Fully open source at github.com/ikchain/Fabrik-Codek (MIT license). Verify the pip package source matches the GitHub repository before installing.