Deterministic agent governance — open-core platform for safe autonomous coding agents.

Website · Docs

---

What is WatchLLM?

WatchLLM is a deterministic runtime governance platform that parses code via Tree-sitter AST, enforces boundary, auth-flow, and secrets rules in real-time, and blocks unsafe writes before they happen.

This repository contains the currently shipped open-core components of WatchLLM — the local kernel, canonical schemas, the VS Code extension source, and example scenarios. A MIT-licensed webproof client SDK is planned but has not shipped; the proving and verification engine remains proprietary.

Architecture

┌──────────────────────────────────────────────────┐
│                  WatchLLM Cloud                  │
│   (Managed platform — closed source)             │
│   ┌──────────┐  ┌──────────┐  ┌───────────────┐  │
│   │ Runtime  │  │  Replay  │  │ Reliability   │  │
│   │ (Rust)   │  │ (DAG UI) │  │ (Adversarial) │  │
│   └──────────┘  └──────────┘  └───────────────┘  │
└──────────────────────────────────────────────────┘
                        │
         Uses schemas & kernel contracts
                        │
┌──────────────────────────────────────────────────┐
│          Open Source (this repo)                 │
│   ┌──────────┐  ┌───────────┐  ┌──────────────┐  │
│   │  Kernel  │  │  Schemas  │  │    VS Code   │  │
│   │ (Python) │  │(Contracts)│  │   Extension  │  │
│   └──────────┘  └───────────┘  └──────────────┘  │
│   ┌───────────────────────────────────────────┐  │
│   │                 Examples                  │  │
│   └───────────────────────────────────────────┘  │
└──────────────────────────────────────────────────┘

Repository Structure

Directory	Description
kernel/	Deterministic Python kernel — Tree-sitter AST parsing, rule engine, CLI
schemas/	Canonical JSON Schema contracts for events, violations, and decisions
vscode/	VS Code extension for save-path governance and inline diagnostics
examples/	Scenario fixtures demonstrating boundary, auth, and secrets rules

Quick Start

Install the kernel

pip install watchllm-kernel

Run a local scan

watchllm-kernel scan ./my-project

Use with VS Code

The extension is not currently published to the VS Code Marketplace. Build and run it locally using the instructions in vscode/README.md.

License

MIT — see LICENSE.

---

WatchLLM Cloud offers the production-grade Rust runtime with DAG replay, adversarial reliability testing, architectural memory (klyd), and multi-model orchestration — all built on the same schemas and kernel contracts in this repo.