Open Source • Local-First • Agentic

Capable.
Controllable.

Lydia is the first local-first AI assistant that treats strategy evolution as a first-class concern. Build trust, eliminate drift.

bash — lydia init
$ npm install -g @lydia-agent/cli
$ lydia init
Initializing repository...Done.
Downloading strategies...v1.0.4 loaded.
$ lydia start
Lydia Runtime started on 127.0.0.1:15536
AI Agentic Loop status: READY
The Challenge

AI Agents Drift.

As agents learn and improve, their behavior often moves away from intended constraints. They cut corners, bypass safety, and become "black boxes" of decision-making.

  • Broad Actions: Agents taking steps beyond their authorized scope.
  • Broken Audits: Hard to understand why an agent changed its strategy.
  • Trust Erosion: Difficulty in predicting behavioral drift over time.
The Solution

Controlled Evolution

Lydia decouples Behavior Strategy from Core Logic. Every behavioral change is a proposal validated through a rigorous safety framework.

Versioned Strategy Files
Offline Replay Evaluation
Risk-Aware Safety Gates

Designed for Trust.

Built from the ground up to be safe, local, and incredibly capable.

Strategy Isolation

Strategies are stored as declarative files. The agent cannot modify its core identity without passing verification.

Branch & Merge

Explore new behaviors in isolated branches. Replay past failures to verify improvements before merging.

Local-First Architecture

Your data never leaves your machine. Full support for local LLMs via Ollama and MCP servers.

Agentic Loop

A transparent "Thinking" process. Monitor every decision, tool call, and observation in real-time.

MCP Native

Extensible with any Model Context Protocol tool. File systems, Git, Web Search, and more.

Human-in-the-Loop

Set risk thresholds that trigger mandatory human review for strategy updates.

Verification

The 4 Safety Gates.

Every strategy proposal must pass through four distinct layers of validation before it becomes active behavior.

01

Syntax & Integrity

Schema validation and structural integrity of the YAML strategy.

02

Evolution Limits

Checking against pre-defined behavioral bounds and risk levels.

03

Offline Replay

Running the new strategy against historical failures (Golden Sets).

04

Human Consensus

Final sign-off for high-impact or ambiguous strategy changes.

Vault Protection

Automated safeguards preventing strategy regression.

Compatibility

Powered by your favorite models.

Anthropic
OpenAI
Ollama
DeepSeek

Build an agent
you can actually trust.