What is Cradle?

Not just an agent — a harness system

Most "local AI agent" tools hand a language model a prompt and forward whatever comes back. Cradle is different: it is a harness system. The language model is the engine, but the value is in everything wrapped around it — an orchestration and control layer that decides which agent answers, grounds the answer in your own documents, gates every reply through a two-tier risk system, and routes anything uncertain to a human operator before it reaches the outside world.

Nothing leaves the perimeter without passing through that harness. That is the whole point: on-premise, offline-capable, auditable AI you can actually put in front of real customers and real data.

What Cradle does

Cradle ingests messages from chat channels (Telegram, an embeddable web widget, or a plain HTTP API), triages each one, drafts a grounded reply with a local LLM, and either sends it automatically or parks it for operator approval:

Incoming message
  → Risk L1 (deterministic rules)     → PII / injection / danger flags
  → Risk L2 (LLM classifier)          → { category, risk, confidence, skills }
  → Router                            → picks the agent by role + skills
  → RAG                               → grounds the draft in your knowledge base
  → Agent                             → drafts a reply with citations
  → Decision:
      green            → auto-reply
      yellow / red     → operator Inbox → approve / edit / reject → send

Every step is recorded: risk assessments, the exact source chunks cited, cost per ticket, and an API audit log. The result is a verified answer with a full paper trail — not an opaque model output.

The pillars of the harness

• Two-tier risk system (СУР). Layer 1 is deterministic rules (drop-table patterns, credentials, PII → red/yellow). Layer 2 is an LLM classifier that returns a structured {category, risk, confidence, reasoning, requiredSkills} verdict. Red and yellow never auto-send.
• Human-in-the-loop approval. Uncertain replies land in an operator Inbox where they can be approved, edited, or rejected before the channel ever sees them.
• Retrieval-augmented grounding. Answers are grounded in your own documents via a local vector search (sqlite-vec) with inline citations — no external API, no data leaving the box.
• Fully local inference. Models run through node-llama-cpp (llama.cpp). After the models are downloaded, Cradle works completely offline. No Ollama, no cloud endpoint required.
• Auditability by construction. Risk log, source citations, cost events and API-key usage are first-class tables, not add-ons.

The three components

Where to go next

• New to Cradle? Install the desktop app.
• Running a GPU box? Follow the GPU server guide.
• Want the mental model? Read the architecture overview.
• Offering Cradle as a service? See Cradle as a datacenter service.
• Using the voice note companion? Read the SmartNotes docs.