GET STARTED
Architecture
How the four services work together on your machine.
System overview
┌─────────────────────────────────────────┐
│ Your Machine │
│ │
│ ┌──────────┐ ┌───────────┐ ┌──────┐ │
│ │ Inference │ │ Embedding │ │ Mesh │ │
│ │ :11435 │ │ :11437 │ │:11436│ │
│ └──────────┘ └───────────┘ └──────┘ │
│ │ │ │ │
│ ┌────────────────────────────────────┐ │
│ │ MCP Server Layer │ │
│ │ tarx-core (29) + tarx-ops (55) │ │
│ │ + tarx-ui (172) + 3 standalone │ │
│ └────────────────────────────────────┘ │
│ │ │
│ ┌──────────┐ │
│ │ SQLite │ memory.db │
│ └──────────┘ │
└─────────────────────────────────────────┘Four services
Each service runs as a separate process on your machine.
llama.cpp server running Qwen 2.5 7B. OpenAI-compatible REST API.
Rust binary for peer-to-peer distributed inference across machines.
llama-server running nomic-embed-text-v1.5. 768-dim vectors.
Higher-order reasoning and planning services.
MCP servers
Six servers provide 307 tools total, accessible from any MCP-compatible client.
Memory, knowledge, project context, system health
Admin, orchestration, automation, audit
End-to-end UI testing across 18 categories
Mesh network health, peers, credits
GTM automation, email, campaigns
Identity verification, evidence, scoring
Storage
All persistent state lives in SQLite — no external databases required.
Memories, files, embeddings (chunk_embeddings + knowledge_embeddings)
Uploaded documents stored on disk
Downloaded model weights
Admin operation log
Path: ~/Library/Application Support/tarx/
Install
Local-first model
TARX installs as a 2.9MB binary (tarxd) via a single curl command. TARX Local is the foundation — it serves inference, manages models, and exposes the MCP endpoint. Everything else builds on top of it.
Knowledge