CRADLE · v0.6

Your agent proposes. Cradle decides what may execute.

Cradle sits between your agent and the systems it can change, on your own hardware. Every tool call arrives as a typed proposal, passes a deterministic gate — contract, permissions, side-effect class, human approval where risk demands it — and lands in a hash-chained log.

One checkpoint between an agent and your systems

Not a chat interface, not a wrapper over a model. This is the path a tool call takes before it becomes an effect.

  1. Agent proposal

    Your agent does not call tools directly. It submits a typed proposal — which tool, which arguments, on whose behalf.

  2. Deterministic gate

    No model involved. Argument contract, tool registration, required scopes, side-effect class from pure to irreversible, idempotency key.

  3. Human approval where required

    Scoped, expiring, single-use, and never self-granted. An unapproved proposal waits instead of executing.

  4. Execution through the gate

    An unregistered tool is not callable. Registration declares what the tool may touch, and which tool compensates it.

  5. Hash-chained audit record

    Proposal, decision, observation — appended with a SHA-256 chain over canonical JSON, so tampering breaks the chain.

Probabilistic — your agent reasonsSymbolic — Cradle governs

What breaks once an agent has real access

A prompt is not a constraint

instruction vs rule

An instruction in a system prompt is a suggestion the model may follow. If it does not, nothing in the system stops the call.

The pause lives inside the agent

approval without limits

Stopping before a tool call is common enough. What is usually missing: a scope for that approval, an expiry, a guarantee it works only once, and a ban on the agent approving its own request.

Traces say how, not why

observability vs decision

A trace tells you which call was made. It does not tell you which rule allowed it, and it does not show whether the record was edited afterwards.

Three steps from install to production

01

Day 1

Install

Cradle Server installs as a single binary on Linux or macOS. Pulls open-weight models. No external dependencies, no telemetry, runs offline from minute one.

02

Days 2-7

Train

Upload corporate documents — PDF, DOCX, XLSX, MD. Optional LoRA fine-tuning on company terminology. Pick from 10+ agent templates: support, legal, sales, HR.

03

Days 8-14

Deploy

Register the tools your agents may call, then reach them over the REST API, an embeddable web widget, Telegram or MCP. Operators monitor draft answers; approve, edit or reject — with full audit log.

What's inside Cradle

The gate is the product. These are the layers it governs, built up around it.

  1. 01Multi-channel — Telegram, embeddable web widget, REST API, MCP
  2. 02EU AI Act compliance — audit log + governance
  3. 03Risk classifier — L1 rules + L2 LLM
  4. 04Agent router by skills and domain
  5. 0510+ agent templates with curated prompts
  6. 06RAG — multi-format parsers (PDF, DOCX, XLSX)
  7. 07sqlite-vec — embedded vector store
  8. 08Web crawler with robots.txt compliance
  9. 09Operator inbox — approve/edit/reject workflow
  10. 10LoRA pipeline — domain fine-tuning
  11. 11Model runner — llama.cpp, open-weight models

Frequently asked questions

Does Cradle ever phone home?

No. After installation it runs entirely offline. The only outbound traffic comes from optional model downloads from Hugging Face — and only when you click "download" on a specific model. You can air-gap the server completely.

What hardware do I need?

For a Mistral 7B-class agent: 32 GB RAM, 16-core CPU, 200 GB SSD. GPU is optional but recommended. Specific guidance for each model is shown in the Models page warning system.

Which models can I run?

Any GGUF-format model: Mistral, Llama 4, Qwen 3, Gemma, Phi, DeepSeek. We curate a recommended list from public benchmarks (OpenLLM Leaderboard, LMSYS Arena), but you can add any custom URL.

How do I integrate with our existing tools?

Cradle exposes a REST API and ships three channels today — Telegram, an embeddable web widget, and the HTTP API itself — plus MCP in both directions, so it can consume your MCP servers and be consumed as one. New channels typically take 200 lines of TypeScript.

Can we fine-tune on our own data?

Yes. Cradle includes a LoRA training pipeline that runs on your hardware (Apple Silicon or NVIDIA GPU). No data leaves your infrastructure during training.

What about EU AI Act compliance?

Cradle ships with audit logging, human-in-the-loop approval and risk classification by default. We provide templates for the documentation Article 11 requires. Final compliance is your responsibility — Cradle makes it possible rather than impossible.

Open source?

Cradle Community (limited) is MIT-licensed; the rest is commercial. The architecture is open — you can read every line of code that runs on your machine.

Where are you based?

Europe. We work remotely with customers across the EU and aligned jurisdictions.

Install it and register your first tool.

The gate runs on your machine. The log stays there too.